input
stringlengths 14
315
| answer
stringlengths 9
2.16k
| gold_ctxs
listlengths 1
15
| ctxs
listlengths 11
186
|
|---|---|---|---|
What are the side effects of group convolution?
|
The side effects of group convolutions are: blocked flow of information between channel groups when multiple group convolutions are combined; and damaged individual convolution filters for each group due to decreased number of input channels [16].
|
[
16
] |
[
{
"id": "1707.01083_all_0",
"text": " Building deeper and larger convolutional neural networks (CNNs) is a primary trend for solving major visual recognition tasks (21, 9, 33, 5, 28, 24). The most accurate CNNs usually have hundreds of layers and thousands of channels (9, 34, 32, 40), thus requiring computation at billions of FLOPs. This report examines the opposite extreme: pursuing the best accuracy in very limited computational budgets at tens or hundreds of MFLOPs, focusing on common mobile platforms such as drones, robots, and smartphones. Note that many existing works (16, 22, 43, 42, 38, 27) focus on pruning, compressing, or low-bit representing a “basic” network architecture. Here we aim to explore a highly efficient basic architecture specially designed for our desired computing ranges. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_1",
"text": " We notice that state-of-the-art basic architectures such as Xception and ResNeXt become less efficient in extremely small networks because of the costly dense 1×1111\\times 1 convolutions. We propose using pointwise group convolutions to reduce computation complexity of 1×1111\\times 1 convolutions. To overcome the side effects brought by group convolutions, we come up with a novel channel shuffle operation to help the information flowing across feature channels. Based on the two techniques, we build a highly efficient architecture called ShuffleNet. Compared with popular structures like (30, 9, 40), for a given computation complexity budget, our ShuffleNet allows more feature map channels, which helps to encode more information and is especially critical to the performance of very small networks. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_2",
"text": " We evaluate our models on the challenging ImageNet classification (4, 29) and MS COCO object detection tasks. A series of controlled experiments shows the effectiveness of our design principles and the better performance over other structures. Compared with the state-of-the-art architecture MobileNet , ShuffleNet achieves superior performance by a significant margin, e.g. absolute 7.8% lower ImageNet top-1 error at level of 40 MFLOPs. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_3",
"text": " We also examine the speedup on real hardware, i.e. an off-the-shelf ARM-based computing core. The ShuffleNet model achieves ∼similar-to\\sim13×\\times actual speedup (theoretical speedup is 18×\\times) over AlexNet while maintaining comparable accuracy. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_4",
"text": " The last few years have seen the success of deep neural networks in computer vision tasks (21, 36, 28), in which model designs play an important role. The increasing needs of running high quality deep neural networks on embedded devices encourage the study on efficient model designs . For example, GoogLeNet increases the depth of networks with much lower complexity compared to simply stacking convolution layers. SqueezeNet reduces parameters and computation significantly while maintaining accuracy. ResNet (9, 10) utilizes the efficient bottleneck structure to achieve impressive performance. SENet introduces an architectural unit that boosts performance at slight computation cost. Concurrent with us, a very recent work employs reinforcement learning and model search to explore efficient model designs. The proposed mobile NASNet model achieves comparable performance with our counterpart ShuffleNet model (26.0% @ 564 MFLOPs vs. 26.3% @ 524 MFLOPs for ImageNet classification error). But do not report results on extremely tiny models (e.g. complexity less than 150 MFLOPs), nor evaluate the actual inference time on mobile devices. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_5",
"text": " The concept of group convolution, which was first introduced in AlexNet for distributing the model over two GPUs, has been well demonstrated its effectiveness in ResNeXt . Depthwise separable convolution proposed in Xception generalizes the ideas of separable convolutions in Inception series (34, 32). Recently, MobileNet utilizes the depthwise separable convolutions and gains state-of-the-art results among lightweight models. Our work generalizes group convolution and depthwise separable convolution in a novel form. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_6",
"text": " To the best of our knowledge, the idea of channel shuffle operation is rarely mentioned in previous work on efficient model design, although CNN library cuda-convnet supports “random sparse convolution” layer, which is equivalent to random channel shuffle followed by a group convolutional layer. Such “random shuffle” operation has different purpose and been seldom exploited later. Very recently, another concurrent work also adopt this idea for a two-stage convolution. However, did not specially investigate the effectiveness of channel shuffle itself and its usage in tiny model design. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_7",
"text": " This direction aims to accelerate inference while preserving accuracy of a pre-trained model. Pruning network connections (6, 7) or channels reduces redundant connections in a pre-trained model while maintaining performance. Quantization (31, 27, 39, 45, 44) and factorization (22, 16, 18, 37) are proposed in literature to reduce redundancy in calculations to speed up inference. Without modifying the parameters, optimized convolution algorithms implemented by FFT (25, 35) and other methods decrease time consumption in practice. Distilling transfers knowledge from large models into small ones, which makes training small models easier. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_8",
"text": " Modern convolutional neural networks (30, 33, 34, 32, 9, 10) usually consist of repeated building blocks with the same structure. Among them, state-of-the-art networks such as Xception and ResNeXt introduce efficient depthwise separable convolutions or group convolutions into the building blocks to strike an excellent trade-off between representation capability and computational cost. However, we notice that both designs do not fully take the 1×1111\\times 1 convolutions (also called pointwise convolutions in ) into account, which require considerable complexity. For example, in ResNeXt only 3×3333\\times 3 layers are equipped with group convolutions. As a result, for each residual unit in ResNeXt the pointwise convolutions occupy 93.4% multiplication-adds (cardinality = 32 as suggested in ). In tiny networks, expensive pointwise convolutions result in limited number of channels to meet the complexity constraint, which might significantly damage the accuracy. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_9",
"text": " To address the issue, a straightforward solution is to apply channel sparse connections, for example group convolutions, also on 1×1111\\times 1 layers. By ensuring that each convolution operates only on the corresponding input channel group, group convolution significantly reduces computation cost. However, if multiple group convolutions stack together, there is one side effect: outputs from a certain channel are only derived from a small fraction of input channels. Fig 1 (a) illustrates a situation of two stacked group convolution layers. It is clear that outputs from a certain group only relate to the inputs within the group. This property blocks information flow between channel groups and weakens representation. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_10",
"text": " If we allow group convolution to obtain input data from different groups (as shown in Fig 1 (b)), the input and output channels will be fully related. Specifically, for the feature map generated from the previous group layer, we can first divide the channels in each group into several subgroups, then feed each group in the next layer with different subgroups. This can be efficiently and elegantly implemented by a channel shuffle operation (Fig 1 (c)): suppose a convolutional layer with g𝑔g groups whose output has g×n𝑔𝑛g\\times n channels; we first reshape the output channel dimension into (g,n)𝑔𝑛(g,n), transposing and then flattening it back as the input of next layer. Note that the operation still takes effect even if the two convolutions have different numbers of groups. Moreover, channel shuffle is also differentiable, which means it can be embedded into network structures for end-to-end training. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_11",
"text": " Channel shuffle operation makes it possible to build more powerful structures with multiple group convolutional layers. In the next subsection we will introduce an efficient network unit with channel shuffle and group convolution. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_12",
"text": " Taking advantage of the channel shuffle operation, we propose a novel ShuffleNet unit specially designed for small networks. We start from the design principle of bottleneck unit in Fig 2 (a). It is a residual block. In its residual branch, for the 3×3333\\times 3 layer, we apply a computational economical 3×3333\\times 3 depthwise convolution on the bottleneck feature map. Then, we replace the first 1×1111\\times 1 layer with pointwise group convolution followed by a channel shuffle operation, to form a ShuffleNet unit, as shown in Fig 2 (b). The purpose of the second pointwise group convolution is to recover the channel dimension to match the shortcut path. For simplicity, we do not apply an extra channel shuffle operation after the second pointwise layer as it results in comparable scores. The usage of batch normalization (BN) and nonlinearity is similar to (9, 40), except that we do not use ReLU after depthwise convolution as suggested by . As for the case where ShuffleNet is applied with stride, we simply make two modifications (see Fig 2 (c)): (i) add a 3×3333\\times 3 average pooling on the shortcut path; (ii) replace the element-wise addition with channel concatenation, which makes it easy to enlarge channel dimension with little extra computation cost. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_13",
"text": " Thanks to pointwise group convolution with channel shuffle, all components in ShuffleNet unit can be computed efficiently. Compared with ResNet (bottleneck design) and ResNeXt , our structure has less complexity under the same settings. For example, given the input size c×h×w𝑐ℎ𝑤c\\times h\\times w and the bottleneck channels m𝑚m, ResNet unit requires hw(2cm+9m2)ℎ𝑤2𝑐𝑚9superscript𝑚2hw(2cm+9m^{2}) FLOPs and ResNeXt has hw(2cm+9m2/g)ℎ𝑤2𝑐𝑚9superscript𝑚2𝑔hw(2cm+9m^{2}/g) FLOPs, while our ShuffleNet unit requires only hw(2cm/g+9m)ℎ𝑤2𝑐𝑚𝑔9𝑚hw(2cm/g+9m) FLOPs, where g𝑔g means the number of groups for convolutions. In other words, given a computational budget, ShuffleNet can use wider feature maps. We find this is critical for small networks, as tiny networks usually have an insufficient number of channels to process the information. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_14",
"text": " In addition, in ShuffleNet depthwise convolution only performs on bottleneck feature maps. Even though depthwise convolution usually has very low theoretical complexity, we find it difficult to efficiently implement on low-power mobile devices, which may result from a worse computation/memory access ratio compared with other dense operations. Such drawback is also referred in , which has a runtime library based on TensorFlow . In ShuffleNet units, we intentionally use depthwise convolution only on bottleneck in order to prevent overhead as much as possible. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_15",
"text": " Built on ShuffleNet units, we present the overall ShuffleNet architecture in Table 1. The proposed network is mainly composed of a stack of ShuffleNet units grouped into three stages. The first building block in each stage is applied with stride = 2. Other hyper-parameters within a stage stay the same, and for the next stage the output channels are doubled. Similar to , we set the number of bottleneck channels to 1/4 of the output channels for each ShuffleNet unit. Our intent is to provide a reference design as simple as possible, although we find that further hyper-parameter tunning might generate better results. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_16",
"text": " In ShuffleNet units, group number g𝑔g controls the connection sparsity of pointwise convolutions. Table 1 explores different group numbers and we adapt the output channels to ensure overall computation cost roughly unchanged (∼similar-to\\sim140 MFLOPs). Obviously, larger group numbers result in more output channels (thus more convolutional filters) for a given complexity constraint, which helps to encode more information, though it might also lead to degradation for an individual convolutional filter due to limited corresponding input channels. In Sec 4.1.1 we will study the impact of this number subject to different computational constrains. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_17",
"text": " To customize the network to a desired complexity, we can simply apply a scale factor s𝑠s on the number of channels. For example, we denote the networks in Table 1 as ”ShuffleNet 1×\\times”, then ”ShuffleNet s×s\\times” means scaling the number of filters in ShuffleNet 1×\\times by s𝑠s times thus overall complexity will be roughly s2superscript𝑠2s^{2} times of ShuffleNet 1×\\times. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_18",
"text": " We mainly evaluate our models on the ImageNet 2012 classification dataset (29, 4). We follow most of the training settings and hyper-parameters used in , with two exceptions: (i) we set the weight decay to 4e-5 instead of 1e-4 and use linear-decay learning rate policy (decreased from 0.5 to 0); (ii) we use slightly less aggressive scale augmentation for data preprocessing. Similar modifications are also referenced in because such small networks usually suffer from underfitting rather than overfitting. It takes 1 or 2 days to train a model for 3×1053superscript1053\\times 10^{5} iterations on 4 GPUs, whose batch size is set to 1024. To benchmark, we compare single crop top-1 performance on ImageNet validation set, i.e. cropping 224×224224224224\\times 224 center view from 256×256\\times input image and evaluating classification accuracy. We use exactly the same settings for all models to ensure fair comparisons. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_19",
"text": " The core idea of ShuffleNet lies in pointwise group convolution and channel shuffle operation. In this subsection we evaluate them respectively. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_20",
"text": " To evaluate the importance of pointwise group convolutions, we compare ShuffleNet models of the same complexity whose numbers of groups range from 1 to 8. If the group number equals 1, no pointwise group convolution is involved and then the ShuffleNet unit becomes an ”Xception-like” structure. For better understanding, we also scale the width of the networks to 3 different complexities and compare their classification performance respectively. Results are shown in Table 2. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_21",
"text": " From the results, we see that models with group convolutions (g>1𝑔1g>1) consistently perform better than the counterparts without pointwise group convolutions (g=1𝑔1g=1). Smaller models tend to benefit more from groups. For example, for ShuffleNet 1×\\times the best entry (g=8𝑔8g=8) is 1.2% better than the counterpart, while for ShuffleNet 0.5×\\times and 0.25×\\times the gaps become 3.5% and 4.4% respectively. Note that group convolution allows more feature map channels for a given complexity constraint, so we hypothesize that the performance gain comes from wider feature maps which help to encode more information. In addition, a smaller network involves thinner feature maps, meaning it benefits more from enlarged feature maps. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_22",
"text": " Table 2 also shows that for some models (e.g. ShuffleNet 0.5×\\times) when group numbers become relatively large (e.g. g=8𝑔8g=8), the classification score saturates or even drops. With an increase in group number (thus wider feature maps), input channels for each convolutional filter become fewer, which may harm representation capability. Interestingly, we also notice that for smaller models such as ShuffleNet 0.25×\\times larger group numbers tend to better results consistently, which suggests wider feature maps bring more benefits for smaller models. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_23",
"text": " The purpose of shuffle operation is to enable cross-group information flow for multiple group convolution layers. Table 3 compares the performance of ShuffleNet structures (group number is set to 3 or 8 for instance) with/without channel shuffle. The evaluations are performed under three different scales of complexity. It is clear that channel shuffle consistently boosts classification scores for different settings. Especially, when group number is relatively large (e.g. g=8𝑔8g=8), models with channel shuffle outperform the counterparts by a significant margin, which shows the importance of cross-group information interchange. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_24",
"text": " Recent leading convolutional units in VGG , ResNet , GoogleNet , ResNeXt and Xception have pursued state-of-the-art results with large models (e.g. ≥1absent1\\geq 1GFLOPs), but do not fully explore low-complexity conditions. In this section we survey a variety of building blocks and make comparisons with ShuffleNet under the same complexity constraint. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_25",
"text": " For fair comparison, we use the overall network architecture as shown in Table 1. We replace the ShuffleNet units in Stage 2-4 with other structures, then adapt the number of channels to ensure the complexity remains unchanged. The structures we explored include: ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_26",
"text": " • VGG-like. Following the design principle of VGG net , we use a two-layer 3×\\times3 convolutions as the basic building block. Different from , we add a Batch Normalization layer after each of the convolutions to make end-to-end training easier. • ResNet. We adopt the ”bottleneck” design in our experiment, which has been demonstrated more efficient in . Same as , the bottleneck ratio111In the bottleneck-like units (like ResNet, ResNeXt or ShuffleNet) bottleneck ratio implies the ratio of bottleneck channels to output channels. For example, bottleneck ratio = 1:4:141:4 means the output feature map is 4 times the width of the bottleneck feature map. is also 1:4:141:4. • Xception-like. The original structure proposed in involves fancy designs or hyper-parameters for different stages, which we find difficult for fair comparison on small models. Instead, we remove the pointwise group convolutions and channel shuffle operation from ShuffleNet (also equivalent to ShuffleNet with g=1𝑔1g=1). The derived structure shares the same idea of “depthwise separable convolution” as in , which is called an Xception-like structure here. • ResNeXt. We use the settings of cardinality =16absent16=16 and bottleneck ratio =1:2:absent12=1:2 as suggested in . We also explore other settings, e.g. bottleneck ratio =1:4:absent14=1:4, and get similar results. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_27",
"text": " We use exactly the same settings to train these models. Results are shown in Table 4. Our ShuffleNet models outperform most others by a significant margin under different complexities. Interestingly, we find an empirical relationship between feature map channels and classification accuracy. For example, under the complexity of 38 MFLOPs, output channels of Stage 4 (see Table 1) for VGG-like, ResNet, ResNeXt, Xception-like, ShuffleNet models are 50, 192, 192, 288, 576 respectively, which is consistent with the increase of accuracy. Since the efficient design of ShuffleNet, we can use more channels for a given computation budget, thus usually resulting in better performance. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_28",
"text": " Note that the above comparisons do not include GoogleNet or Inception series (33, 34, 32). We find it nontrivial to generate such Inception structures to small networks because the original design of Inception module involves too many hyper-parameters. As a reference, the first GoogleNet version has 31.3% top-1 error at the cost of 1.5 GFLOPs (See Table 6). More sophisticated Inception versions (34, 32) are more accurate, however, involve significantly increased complexity. Recently, Kim et al. propose a lightweight network structure named PVANET which adopts Inception units. Our reimplemented PVANET (with 224×\\times224 input size) has 29.7% classification error with a computation complexity of 557 MFLOPs, while our ShuffleNet 2x model (g=3𝑔3g=3) gets 26.3% with 524 MFLOPs (see Table 6). ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_29",
"text": " Recently Howard et al. have proposed MobileNets which mainly focus on efficient network architecture for mobile devices. MobileNet takes the idea of depthwise separable convolution from and achieves state-of-the-art results on small models. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_30",
"text": " Table 5 compares classification scores under a variety of complexity levels. It is clear that our ShuffleNet models are superior to MobileNet for all the complexities. Though our ShuffleNet network is specially designed for small models (<150absent150<150 MFLOPs), we find it is still better than MobileNet for higher computation cost, e.g. 3.1% more accurate than MobileNet 1×\\times at the cost of 500 MFLOPs. For smaller networks (∼similar-to\\sim40 MFLOPs) ShuffleNet surpasses MobileNet by 7.8%. Note that our ShuffleNet architecture contains 50 layers while MobileNet only has 28 layers. For better understanding, we also try ShuffleNet on a 26-layer architecture by removing half of the blocks in Stage 2-4 (see ”ShuffleNet 0.5×\\times shallow (g=3𝑔3g=3)” in Table 5). Results show that the shallower model is still significantly better than the corresponding MobileNet, which implies that the effectiveness of ShuffleNet mainly results from its efficient structure, not the depth. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_31",
"text": " Table 6 compares our ShuffleNet with a few popular models. Results show that with similar accuracy ShuffleNet is much more efficient than others. For example, ShuffleNet 0.5×\\times is theoretically 18×\\times faster than AlexNet with comparable classification score. We will evaluate the actual running time in Sec 4.5. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_32",
"text": " It is also worth noting that the simple architecture design makes it easy to equip ShuffeNets with the latest advances such as (13, 26). For example, in the authors propose Squeeze-and-Excitation (SE) blocks which achieve state-of-the-art results on large ImageNet models. We find SE modules also take effect in combination with the backbone ShuffleNets, for instance, boosting the top-1 error of ShuffleNet 2×\\times to 24.7% (shown in Table 5). Interestingly, though negligible increase of theoretical complexity, we find ShuffleNets with SE modules are usually 25∼40%similar-to25percent4025\\sim 40\\% slower than the “raw” ShuffleNets on mobile devices, which implies that actual speedup evaluation is critical on low-cost architecture design. In Sec 4.5 we will make further discussion. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_33",
"text": " To evaluate the generalization ability for transfer learning, we test our ShuffleNet model on the task of MS COCO object detection . We adopt Faster-RCNN as the detection framework and use the publicly released Caffe code (28, 17) for training with default settings. Similar to , the models are trained on the COCO train+val dataset excluding 5000 minival images and we conduct testing on the minival set. Table 7 shows the comparison of results trained and evaluated on two input resolutions. Comparing ShuffleNet 2×\\times with MobileNet whose complexity are comparable (524 vs. 569 MFLOPs), our ShuffleNet 2×\\times surpasses MobileNet by a significant margin on both resolutions; our ShuffleNet 1×\\times also achieves comparable results with MobileNet on 600×\\times resolution, but has ∼similar-to\\sim4×\\times complexity reduction. We conjecture that this significant gain is partly due to ShuffleNet’s simple design of architecture without bells and whistles. ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
},
{
"id": "1707.01083_all_34",
"text": " Finally, we evaluate the actual inference speed of ShuffleNet models on a mobile device with an ARM platform. Though ShuffleNets with larger group numbers (e.g. g=4𝑔4g=4 or g=8𝑔8g=8) usually have better performance, we find it less efficient in our current implementation. Empirically g=3𝑔3g=3 usually has a proper trade-off between accuracy and actual inference time. As shown in Table 8, three input resolutions are exploited for the test. Due to memory access and other overheads, we find every 4×\\times theoretical complexity reduction usually results in ∼similar-to\\sim2.6×\\times actual speedup in our implementation. Nevertheless, compared with AlexNet our ShuffleNet 0.5×\\times model still achieves ∼similar-to\\sim13×\\times actual speedup under comparable classification accuracy (the theoretical speedup is 18×\\times), which is much faster than previous AlexNet-level models or speedup approaches such as (14, 16, 22, 42, 43, 38). ",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
}
] |
How does normalizing the features and weights in the softmax loss function improve the performance of deep face recognition systems?
|
Normalizing the weights only can help angular/cosine-margin-based loss to make the learned features more discriminative, whereas normalizing only the learned features can help overcome the bias to the sample distribution of the softmax [25]. Since L2-norms of learned features with softmax loss were observed to be reflective of the quality of the face, making all the features have the same L2-norm may help to give similar attention to all different qualities of samples [26].
|
[
25,
26
] |
[
{
"id": "1804.06655_all_0",
"text": " Face recognition (FR) has been the prominent biometric technique for identity authentication and has been widely used in many areas, such as military, finance, public security and daily life. FR has been a long-standing research topic in the CVPR community. In the early 1990s, the study of FR became popular following the introduction of the historical Eigenface approach . The milestones of feature-based FR over the past years are presented in Fig. 1, in which the times of four major technical streams are highlighted. The holistic approaches derive the low-dimensional representation through certain distribution assumptions, such as linear subspace , manifold , and sparse representation . This idea dominated the FR community in the 1990s and 2000s. However, a well-known problem is that these theoretically plausible holistic methods fail to address the uncontrolled facial changes that deviate from their prior assumptions. In the early 2000s, this problem gave rise to local-feature-based FR. Gabor and LBP , as well as their multilevel and high-dimensional extensions , achieved robust performance through some invariant properties of local filtering. Unfortunately, handcrafted features suffered from a lack of distinctiveness and compactness. In the early 2010s, learning-based local descriptors were introduced to the FR community , in which local filters are learned for better distinctiveness and the encoding codebook is learned for better compactness. However, these shallow representations still have an inevitable limitation on robustness against the complex nonlinear facial appearance variations. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_1",
"text": " In general, traditional methods attempted to recognize human face by one or two layer representations, such as filtering responses, histogram of the feature codes, or distribution of the dictionary atoms. The research community studied intensively to separately improve the preprocessing, local descriptors, and feature transformation, but these approaches improved FR accuracy slowly. What’s worse, most methods aimed to address one aspect of unconstrained facial changes only, such as lighting, pose, expression, or disguise. There was no any integrated technique to address these unconstrained challenges integrally. As a result, with continuous efforts of more than a decade, “shallow” methods only improved the accuracy of the LFW benchmark to about 95% , which indicates that “shallow” methods are insufficient to extract stable identity feature invariant to real-world changes. Due to the insufficiency of this technical, facial recognition systems were often reported with unstable performance or failures with countless false alarms in real-world applications. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_2",
"text": " But all that changed in 2012 when AlexNet won the ImageNet competition by a large margin using a technique called deep learning . Deep learning methods, such as convolutional neural networks, use a cascade of multiple layers of processing units for feature extraction and transformation. They learn multiple levels of representations that correspond to different levels of abstraction. The levels form a hierarchy of concepts, showing strong invariance to the face pose, lighting, and expression changes, as shown in Fig. 2. It can be seen from the figure that the first layer of the deep neural network is somewhat similar to the Gabor feature found by human scientists with years of experience. The second layer learns more complex texture features. The features of the third layer are more complex, and some simple structures have begun to appear, such as high-bridged nose and big eyes. In the fourth, the network output is enough to explain a certain facial attribute, which can make a special response to some clear abstract concepts such as smile, roar, and even blue eye. In conclusion, in deep convolutional neural networks (CNN), the lower layers automatically learn the features similar to Gabor and SIFT designed for years or even decades (such as initial layers in Fig. 2), and the higher layers further learn higher level abstraction. Finally, the combination of these higher level abstraction represents facial identity with unprecedented stability. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_3",
"text": " In 2014, DeepFace achieved the SOTA accuracy on the famous LFW benchmark , approaching human performance on the unconstrained condition for the first time (DeepFace: 97.35% vs. Human: 97.53%), by training a 9-layer model on 4 million facial images. Inspired by this work, research focus has shifted to deep-learning-based approaches, and the accuracy was dramatically boosted to above 99.80% in just three years. Deep learning technique has reshaped the research landscape of FR in almost all aspects such as algorithm designs, training/test datasets, application scenarios and even the evaluation protocols. Therefore, it is of great significance to review the breakthrough and rapid development process in recent years. There have been several surveys on FR (24, 25, 26, 27, 28) and its subdomains, and they mostly summarized and compared a diverse set of techniques related to a specific FR scene, such as illumination-invariant FR , 3D FR , pose-invariant FR . Unfortunately, due to their earlier publication dates, none of them covered the deep learning methodology that is most successful nowadays. This survey focuses only on recognition problem, and one can refer to Ranjan et al. for a brief review of a full deep FR pipeline with detection and alignment, or refer to Jin et al. for a survey of face alignment. Specifically, the major contributions of this survey are as follows: ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_4",
"text": " • A systematic review on the evolution of the network architectures and loss functions for deep FR is provided. Various loss functions are categorized into Euclidean-distance-based loss, angular/cosine-margin-based loss and softmax loss and its variations. Both the mainstream network architectures, such as Deepface , DeepID series (34, 35, 21, 36), VGGFace , FaceNet , and VGGFace2 , and other architectures designed for FR are covered. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_5",
"text": " • We categorize the new face processing methods based on deep learning, such as those used to handle recognition difficulty on pose changes, into two classes: “one-to-many augmentation” and “many-to-one normalization”, and discuss how emerging generative adversarial network (GAN) facilitates deep FR. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_6",
"text": " • We present a comparison and analysis on public available databases that are of vital importance for both model training and testing. Major FR benchmarks, such as LFW , IJB-A/B/C (41, 42, 43), Megaface , and MS-Celeb-1M , are reviewed and compared, in term of the four aspects: training methodology, evaluation tasks and metrics, and recognition scenes, which provides an useful reference for training and testing deep FR. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_7",
"text": " • Besides the general purpose tasks defined by the major databases, we summarize a dozen scenario-specific databases and solutions that are still challenging for deep learning, such as anti-attack, cross-pose FR, and cross-age FR. By reviewing specially designed methods for these unsolved problems, we attempt to reveal the important issues for future research on deep FR, such as adversarial samples, algorithm/data biases, and model interpretability. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_8",
"text": " The remainder of this survey is structured as follows. In Section II, we introduce some background concepts and terminologies, and then we briefly introduce each component of FR. In Section III, different network architectures and loss functions are presented. Then, we summarize the face processing algorithms and the datasets. In Section V, we briefly introduce several methods of deep FR used for different scenes. Finally, the conclusion of this paper and discussion of future works are presented in Section VI. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_9",
"text": " As mentioned in , there are three modules needed for FR system, as shown in Fig. 3. First, a face detector is used to localize faces in images or videos. Second, with the facial landmark detector, the faces are aligned to normalized canonical coordinates. Third, the FR module is implemented with these aligned face images. We only focus on the FR module throughout the remainder of this paper. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_10",
"text": " Before a face image is fed to an FR module, face anti-spoofing, which recognizes whether the face is live or spoofed, is applied to avoid different types of attacks. Then, recognition can be performed. As shown in Fig. 3(c), an FR module consists of face processing, deep feature extraction and face matching, and it can be described as follows: ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_11",
"text": " M(F(Pi(Ii)),F(Pj(Ij)))𝑀𝐹subscript𝑃𝑖subscript𝐼𝑖𝐹subscript𝑃𝑗subscript𝐼𝑗M(F(P_{i}(I_{i})),F(P_{j}(I_{j}))) (1) where Iisubscript𝐼𝑖I_{i} and Ijsubscript𝐼𝑗I_{j} are two face images, respectively. P𝑃P stands for face processing to handle intra-personal variations before training and testing, such as poses, illuminations, expressions and occlusions. F𝐹F denotes feature extraction, which encodes the identity information. The feature extractor is learned by loss functions when training, and is utilized to extract features of faces when testing. M𝑀M means a face matching algorithm used to compute similarity scores of features to determine the specific identity of faces. Different from object classification, the testing identities are usually disjoint from the training data in FR, which makes the learned classifier cannot be used to recognize testing faces. Therefore, face matching algorithm is an essential part in FR. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_12",
"text": " Although deep-learning-based approaches have been widely used, Mehdipour et al. proved that various conditions, such as poses, illuminations, expressions and occlusions, still affect the performance of deep FR. Accordingly, face processing is introduced to address this problem. The face processing methods are categorized as “one-to-many augmentation” and “many-to-one normalization”, as shown in Table I. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_13",
"text": " • “One-to-many augmentation”. These methods generate many patches or images of the pose variability from a single image to enable deep networks to learn pose-invariant representations. • “Many-to-one normalization”. These methods recover the canonical view of face images from one or many images of a nonfrontal view; then, FR can be performed as if it were under controlled conditions. Note that we mainly focus on deep face processing method designed for pose variations in this paper, since pose is widely regarded as a major challenge in automatic FR applications and other variations can be solved by the similar methods. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_14",
"text": " Network Architecture. The architectures can be categorized as backbone and assembled networks, as shown in Table II. Inspired by the extraordinary success on the ImageNet challenge, the typical CNN architectures, e.g. AlexNet, VGGNet, GoogleNet, ResNet and SENet (22, 75, 76, 77, 78), are introduced and widely used as the baseline models in FR (directly or slightly modified). In addition to the mainstream, some assembled networks, e.g. multi-task networks and multi-input networks, are utilized in FR. Hu et al. shows that accumulating the results of assembled networks provides an increase in performance compared with an individual network. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_15",
"text": " Loss Function. The softmax loss is commonly used as the supervision signal in object recognition, and it encourages the separability of features. However, the softmax loss is not sufficiently effective for FR because intra-variations could be larger than inter-differences and more discriminative features are required when recognizing different people. Many works focus on creating novel loss functions to make features not only more separable but also discriminative, as shown in Table III. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_16",
"text": " FR can be categorized as face verification and face identification. In either scenario, a set of known subjects is initially enrolled in the system (the gallery), and during testing, a new subject (the probe) is presented. After the deep networks are trained on massive data with the supervision of an appropriate loss function, each of the test images is passed through the networks to obtain a deep feature representation. Using cosine distance or L2 distance, face verification computes one-to-one similarity between the gallery and probe to determine whether the two images are of the same subject, whereas face identification computes one-to-many similarity to determine the specific identity of a probe face. In addition to these, other methods are introduced to postprocess the deep features such that the face matching is performed efficiently and accurately, such as metric learning, sparse-representation-based classifier (SRC), and so forth. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_17",
"text": " To sum up, we present FR modules and their commonly-used methods in Fig. 4 to help readers to get a view of the whole FR. In deep FR, various training and testing face databases are constructed, and different architectures and losses of deep FR always follow those of deep object classification and are modified according to unique characteristics of FR. Moreover, in order to address unconstrained facial changes, face processing methods are further designed to handle poses, expressions and occlusions variations. Benefiting from these strategies, deep FR system significantly improves the SOTA and surpasses human performance. When the applications of FR becomes more and more mature in general scenario, recently, different solutions are driven for more difficult specific scenarios, such as cross-pose FR, cross-age FR, video FR. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_18",
"text": " For most applications, it is difficult to include the candidate faces during the training stage, which makes FR become a “zero-shot” learning task. Fortunately, since all human faces share a similar shape and texture, the representation learned from a small proportion of faces can generalize well to the rest. Based on this theory, a straightforward way to improve generalized performance is to include as many IDs as possible in the training set. For example, Internet giants such as Facebook and Google have reported their deep FR system trained by 106−107superscript106superscript10710^{6}-10^{7} IDs (38, 20). ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_19",
"text": " Unfortunately, these personal datasets, as well as prerequisite GPU clusters for distributed model training, are not accessible for academic community. Currently, public available training databases for academic research consist of only 103−105superscript103superscript10510^{3}-10^{5} IDs. Instead, academic community makes effort to design effective loss functions and adopts efficient architectures to make deep features more discriminative using the relatively small training data sets. For instance, the accuracy of most popular LFW benchmark has been boosted from 97% to above 99.8% in the pasting four years, as enumerated in Table IV. In this section, we survey the research efforts on different loss functions and network architectures that have significantly improved deep FR methods. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_20",
"text": " Inheriting from the object classification network such as AlexNet, the initial Deepface and DeepID adopted cross-entropy based softmax loss for feature learning. After that, people realized that the softmax loss is not sufficient by itself to learn discriminative features, and more researchers began to explore novel loss functions for enhanced generalization ability. This becomes the hottest research topic in deep FR research, as illustrated in Fig. 5. Before 2017, Euclidean-distance-based loss played an important role; In 2017, angular/cosine-margin-based loss as well as feature and weight normalization became popular. It should be noted that, although some loss functions share the similar basic idea, the new one is usually designed to facilitate the training procedure by easier parameter or sample selection. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_21",
"text": " Euclidean-distance-based loss is a metric learning method (118, 119) that embeds images into Euclidean space in which intra-variance is reduced and inter-variance is enlarged. The contrastive loss and the triplet loss are the commonly used loss functions. The contrastive loss (35, 21, 36, 61, 120) requires face image pairs, and then pulls together positive pairs and pushes apart negative pairs. ℒ=yijmax(0,‖f(xi)−f(xj)‖2−ϵ+)+(1−yij)max(0,ϵ−−‖f(xi)−f(xj)‖2)ℒsubscript𝑦𝑖𝑗𝑚𝑎𝑥0subscriptdelimited-∥∥𝑓subscript𝑥𝑖𝑓subscript𝑥𝑗2superscriptitalic-ϵ1subscript𝑦𝑖𝑗𝑚𝑎𝑥0superscriptitalic-ϵsubscriptdelimited-∥∥𝑓subscript𝑥𝑖𝑓subscript𝑥𝑗2\\begin{split}\\mathcal{L}=&y_{ij}max\\left(0,\\left\\|f(x_{i})-f(x_{j})\\right\\|_{2}-\\epsilon^{+}\\right)\\\\ &+(1-y_{ij})max\\left(0,\\epsilon^{-}-\\left\\|f(x_{i})-f(x_{j})\\right\\|_{2}\\right)\\end{split} (2) where yij=1subscript𝑦𝑖𝑗1y_{ij}=1 means xisubscript𝑥𝑖x_{i} and xjsubscript𝑥𝑗x_{j} are matching samples and yij=0subscript𝑦𝑖𝑗0y_{ij}=0 means non-matching samples. f(⋅)𝑓⋅f(\\cdot) is the feature embedding, ϵ+superscriptitalic-ϵ\\epsilon^{+} and ϵ−superscriptitalic-ϵ\\epsilon^{-} control the margins of the matching and non-matching pairs respectively. DeepID2 combined the face identification (softmax) and verification (contrastive loss) supervisory signals to learn a discriminative representation, and joint Bayesian (JB) was applied to obtain a robust embedding space. Extending from DeepID2 , DeepID2+ increased the dimension of hidden representations and added supervision to early convolutional layers. DeepID3 further introduced VGGNet and GoogleNet to their work. However, the main problem with the contrastive loss is that the margin parameters are often difficult to choose. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_22",
"text": " Contrary to contrastive loss that considers the absolute distances of the matching pairs and non-matching pairs, triplet loss considers the relative difference of the distances between them. Along with FaceNet proposed by Google, Triplet loss (38, 37, 81, 80, 58, 60) was introduced into FR. It requires the face triplets, and then it minimizes the distance between an anchor and a positive sample of the same identity and maximizes the distance between the anchor and a negative sample of a different identity. FaceNet made ‖f(xia)−f(xip)‖22+α<−‖f(xia)−f(xin)‖22superscriptsubscriptnorm𝑓superscriptsubscript𝑥𝑖𝑎𝑓superscriptsubscript𝑥𝑖𝑝22𝛼superscriptsubscriptnorm𝑓superscriptsubscript𝑥𝑖𝑎𝑓superscriptsubscript𝑥𝑖𝑛22\\left\\|f(x_{i}^{a})-f(x_{i}^{p})\\right\\|_{2}^{2}+\\alpha<-\\left\\|f(x_{i}^{a})-f(x_{i}^{n})\\right\\|_{2}^{2} using hard triplet face samples, where xiasuperscriptsubscript𝑥𝑖𝑎x_{i}^{a}, xipsuperscriptsubscript𝑥𝑖𝑝x_{i}^{p} and xinsuperscriptsubscript𝑥𝑖𝑛x_{i}^{n} are the anchor, positive and negative samples, respectively, α𝛼\\alpha is a margin and f(⋅)𝑓⋅f(\\cdot) represents a nonlinear transformation embedding an image into a feature space. Inspired by FaceNet , TPE and TSE learned a linear projection W𝑊W to construct triplet loss. Other methods optimize deep models using both triplet loss and softmax loss (59, 58, 60, 121). They first train networks with softmax and then fine-tune them with triplet loss. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_23",
"text": " However, the contrastive loss and triplet loss occasionally encounter training instability due to the selection of effective training samples, some paper begun to explore simple alternatives. Center loss and its variants (82, 116, 102) are good choices for reducing intra-variance. The center loss learned a center for each class and penalized the distances between the deep features and their corresponding class centers. This loss can be defined as follows: ℒC=12∑i=1m‖xi−cyi‖22subscriptℒ𝐶12superscriptsubscript𝑖1𝑚superscriptsubscriptnormsubscript𝑥𝑖subscript𝑐subscript𝑦𝑖22\\mathcal{L}_{C}=\\frac{1}{2}\\sum_{i=1}^{m}\\left\\|x_{i}-c_{y_{i}}\\right\\|_{2}^{2} (3) where xisubscript𝑥𝑖x_{i} denotes the i𝑖i-th deep feature belonging to the yisubscript𝑦𝑖y_{i}-th class and cyisubscript𝑐subscript𝑦𝑖c_{y_{i}} denotes the yisubscript𝑦𝑖y_{i}-th class center of deep features. To handle the long-tailed data, a range loss , which is a variant of center loss, is used to minimize k greatest range’s harmonic mean values in one class and maximize the shortest inter-class distance within one batch. Wu et al. proposed a center-invariant loss that penalizes the difference between each center of classes. Deng et al. selected the farthest intra-class samples and the nearest inter-class samples to compute a margin loss. However, the center loss and its variants suffer from massive GPU memory consumption on the classification layer, and prefer balanced and sufficient training data for each identity. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_24",
"text": " In 2017, people had a deeper understanding of loss function in deep FR and thought that samples should be separated more strictly to avoid misclassifying the difficult samples. Angular/cosine-margin-based loss (104, 84, 105, 106, 108) is proposed to make learned features potentially separable with a larger angular/cosine distance. The decision boundary in softmax loss is (W1−W2)x+b1−b2=0subscript𝑊1subscript𝑊2𝑥subscript𝑏1subscript𝑏20\\left(W_{1}-W_{2}\\right)x+b_{1}-b_{2}=0, where x𝑥x is feature vector, Wisubscript𝑊𝑖W_{i} and bisubscript𝑏𝑖b_{i} are weights and bias in softmax loss, respectively. Liu et al. reformulated the original softmax loss into a large-margin softmax (L-Softmax) loss. They constrain b1=b2=0subscript𝑏1subscript𝑏20b_{1}=b_{2}=0, so the decision boundaries for class 1 and class 2 become ‖x‖(‖W1‖cos(mθ1)−‖W2‖cos(θ2))=0norm𝑥normsubscript𝑊1𝑐𝑜𝑠𝑚subscript𝜃1normsubscript𝑊2𝑐𝑜𝑠subscript𝜃20\\left\\|x\\right\\|\\left(\\left\\|W_{1}\\right\\|cos\\left(m\\theta_{1}\\right)-\\left\\|W_{2}\\right\\|cos\\left(\\theta_{2}\\right)\\right)=0 and ‖x‖(‖W1‖‖W2‖cos(θ1)−cos(mθ2))=0norm𝑥normsubscript𝑊1normsubscript𝑊2𝑐𝑜𝑠subscript𝜃1𝑐𝑜𝑠𝑚subscript𝜃20\\left\\|x\\right\\|\\left(\\left\\|W_{1}\\right\\|\\left\\|W_{2}\\right\\|cos\\left(\\theta_{1}\\right)-cos\\left(m\\theta_{2}\\right)\\right)=0, respectively, where m𝑚m is a positive integer introducing an angular margin, and θisubscript𝜃𝑖\\theta_{i} is the angle between Wisubscript𝑊𝑖W_{i} and x𝑥x. Due to the non-monotonicity of the cosine function, a piece-wise function is applied in L-softmax to guarantee the monotonicity. The loss function is defined as follows: ℒi=−log(e‖Wyi‖‖xi‖φ(θyi)e‖Wyi‖‖xi‖φ(θyi)+∑j≠yie‖Wyi‖‖xi‖cos(θj))subscriptℒ𝑖𝑙𝑜𝑔superscript𝑒normsubscript𝑊𝑦𝑖normsubscript𝑥𝑖𝜑subscript𝜃𝑦𝑖superscript𝑒normsubscript𝑊𝑦𝑖normsubscript𝑥𝑖𝜑subscript𝜃𝑦𝑖subscript𝑗subscript𝑦𝑖superscript𝑒normsubscript𝑊𝑦𝑖normsubscript𝑥𝑖𝑐𝑜𝑠subscript𝜃𝑗\\mathcal{L}_{i}=-log\\left(\\frac{e^{\\left\\|W_{yi}\\right\\|\\left\\|x_{i}\\right\\|\\varphi(\\theta_{yi})}}{e^{\\left\\|W_{yi}\\right\\|\\left\\|x_{i}\\right\\|\\varphi(\\theta_{yi})+\\sum_{j\\neq y_{i}}e^{\\left\\|W_{yi}\\right\\|\\left\\|x_{i}\\right\\|cos(\\theta_{j})}}}\\right) (4) where φ(θ)=(−1)kcos(mθ)−2k,θ∈(kπm,(k+1)πm)formulae-sequence𝜑𝜃superscript1𝑘𝑐𝑜𝑠𝑚𝜃2𝑘𝜃𝑘𝜋𝑚𝑘1𝜋𝑚\\varphi(\\theta)=(-1)^{k}cos(m\\theta)-2k,\\theta\\in\\left(\\frac{k\\pi}{m},\\frac{(k+1)\\pi}{m}\\right) (5) Considering that L-Softmax is difficult to converge, it is always combined with softmax loss to facilitate and ensure the convergence. Therefore, the loss function is changed into: fyi=λ‖Wyi‖‖xi‖cos(θyi)+‖Wyi‖‖xi‖φ(θyi)1+λsubscript𝑓subscript𝑦𝑖𝜆normsubscript𝑊subscript𝑦𝑖normsubscript𝑥𝑖𝑐𝑜𝑠subscript𝜃subscript𝑦𝑖normsubscript𝑊subscript𝑦𝑖normsubscript𝑥𝑖𝜑subscript𝜃subscript𝑦𝑖1𝜆f_{y_{i}}=\\frac{\\lambda\\left\\|W_{y_{i}}\\right\\|\\left\\|x_{i}\\right\\|cos(\\theta_{y_{i}})+\\left\\|W_{y_{i}}\\right\\|\\left\\|x_{i}\\right\\|\\varphi(\\theta_{y_{i}})}{1+\\lambda}, where λ𝜆\\lambda is a dynamic hyper-parameter. Based on L-Softmax, A-Softmax loss further normalized the weight W𝑊W by L2 norm (‖W‖=1norm𝑊1\\left\\|W\\right\\|=1) such that the normalized vector will lie on a hypersphere, and then the discriminative face features can be learned on a hypersphere manifold with an angular margin (Fig. 6). Liu et al. introduced a deep hyperspherical convolution network (SphereNet) that adopts hyperspherical convolution as its basic convolution operator and is supervised by angular-margin-based loss. To overcome the optimization difficulty of L-Softmax and A-Softmax, which incorporate the angular margin in a multiplicative manner, ArcFace and CosFace , AMS loss respectively introduced an additive angular/cosine margin cos(θ+m)𝑐𝑜𝑠𝜃𝑚cos(\\theta+m) and cosθ−m𝑐𝑜𝑠𝜃𝑚cos\\theta-m. They are extremely easy to implement without tricky hyper-parameters λ𝜆\\lambda, and are more clear and able to converge without the softmax supervision. The decision boundaries under the binary classification case are given in Table V. Based on large margin, FairLoss and AdaptiveFace further proposed to adjust the margins for different classes adaptively to address the problem of unbalanced data. Compared to Euclidean-distance-based loss, angular/cosine-margin-based loss explicitly adds discriminative constraints on a hypershpere manifold, which intrinsically matches the prior that human face lies on a manifold. However, Wang et al. showed that angular/cosine-margin-based loss can achieve better results on a clean dataset, but is vulnerable to noise and becomes worse than center loss and softmax in the high-noise region as shown in Fig. 7. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_25",
"text": " In 2017, in addition to reformulating softmax loss into an angular/cosine-margin-based loss as mentioned above, some works tries to normalize the features and weights in loss functions to improve the model performance, which can be written as follows: W^=W‖W‖,x^=αx‖x‖formulae-sequence^𝑊𝑊norm𝑊^𝑥𝛼𝑥norm𝑥\\hat{W}=\\frac{W}{\\left\\|W\\right\\|},\\hat{x}=\\alpha\\frac{x}{\\left\\|x\\right\\|} (6) where α𝛼\\alpha is a scaling parameter, x𝑥x is the learned feature vector, W𝑊W is weight of last fully connected layer. Scaling x𝑥x to a fixed radius α𝛼\\alpha is important, as Wang et al. proved that normalizing both features and weights to 1 will make the softmax loss become trapped at a very high value on the training set. After that, the loss function, e.g. softmax, can be performed using the normalized features and weights. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_26",
"text": " Some papers (84, 108) first normalized the weights only and then added angular/cosine margin into loss functions to make the learned features be discriminative. In contrast, some works, such as (109, 111), adopted feature normalization only to overcome the bias to the sample distribution of the softmax. Based on the observation of that the L2-norm of features learned using the softmax loss is informative of the quality of the face, L2-softmax enforced all the features to have the same L2-norm by feature normalization such that similar attention is given to good quality frontal faces and blurry faces with extreme pose. Rather than scaling x𝑥x to the parameter α𝛼\\alpha, Hasnat et al. normalized features with x^=x−μσ2^𝑥𝑥𝜇superscript𝜎2\\hat{x}=\\frac{x-\\mu}{\\sqrt{\\sigma^{2}}}, where μ𝜇\\mu and σ2superscript𝜎2\\sigma^{2} are the mean and variance. Ring loss encouraged the norm of samples being value R𝑅R (a learned parameter) rather than explicit enforcing through a hard normalization operation. Moreover, normalizing both features and weights (110, 112, 115, 105, 106) has become a common strategy. Wang et al. explained the necessity of this normalization operation from both analytic and geometric perspectives. After normalizing features and weights, CoCo loss optimized the cosine distance among data features, and Hasnat et al. used the von Mises-Fisher (vMF) mixture model as the theoretical basis to develop a novel vMF mixture loss and its corresponding vMF deep features. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_27",
"text": " Mainstream architectures. The commonly used network architectures of deep FR have always followed those of deep object classification and evolved from AlexNet to SENet rapidly. We present the most influential architectures of deep object classification and deep face recognition in chronological order 111The time we present is when the paper was published. in Fig. 8. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_28",
"text": " In 2012, AlexNet was reported to achieve the SOTA recognition accuracy in the ImageNet large-scale visual recognition competition (ILSVRC) 2012, exceeding the previous best results by a large margin. AlexNet consists of five convolutional layers and three fully connected layers, and it also integrates various techniques, such as rectified linear unit (ReLU), dropout, data augmentation, and so forth. ReLU was widely regarded as the most essential component for making deep learning possible. Then, in 2014, VGGNet proposed a standard network architecture that used very small 3×3333\\times 3 convolutional filters throughout and doubled the number of feature maps after the 2×\\times2 pooling. It increased the depth of the network to 16-19 weight layers, which further enhanced the flexibility to learn progressive nonlinear mappings by deep architectures. In 2015, the 22-layer GoogleNet introduced an “inception module” with the concatenation of hybrid feature maps, as well as two additional intermediate softmax supervised signals. It performs several convolutions with different receptive fields (1×1111\\times 1, 3×3333\\times 3 and 5×5555\\times 5) in parallel, and concatenates all feature maps to merge the multi-resolution information. In 2016, ResNet proposed to make layers learn a residual mapping with reference to the layer inputs ℱ(x):=ℋ(x)−xassignℱ𝑥ℋ𝑥𝑥\\mathcal{F}(x):=\\mathcal{H}(x)-x rather than directly learning a desired underlying mapping ℋ(x)ℋ𝑥\\mathcal{H}(x) to ease the training of very deep networks (up to 152 layers). The original mapping is recast into ℱ(x)+xℱ𝑥𝑥\\mathcal{F}(x)+x and can be realized by “shortcut connections”. As the champion of ILSVRC 2017, SENet introduced a “Squeeze-and-Excitation” (SE) block, that adaptively recalibrates channel-wise feature responses by explicitly modelling interdependencies between channels. These blocks can be integrated with modern architectures, such as ResNet, and improves their representational power. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_29",
"text": " With the evolved architectures and advanced training techniques, such as batch normalization (BN), the network becomes deeper and the training becomes more controllable. Following these architectures in object classification, the networks in deep FR are also developed step by step, and the performance of deep FR is continually improving. We present these mainstream architectures of deep FR in Fig. 9. In 2014, DeepFace was the first to use a nine-layer CNN with several locally connected layers. With 3D alignment for face processing, it reaches an accuracy of 97.35% on LFW. In 2015, FaceNet used a large private dataset to train a GoogleNet. It adopted a triplet loss function based on triplets of roughly aligned matching/nonmatching face patches generated by a novel online triplet mining method and achieved good performance of 99.63%. In the same year, VGGface designed a procedure to collect a large-scale dataset from the Internet. It trained the VGGNet on this dataset and then fine-tuned the networks via a triplet loss function similar to FaceNet. VGGface obtains an accuracy of 98.95%. In 2017, SphereFace used a 64-layer ResNet architecture and proposed the angular softmax (A-Softmax) loss to learn discriminative face features with angular margin. It boosts the achieves to 99.42% on LFW. In the end of 2017, a new large-scale face dataset, namely VGGface2 , was introduced, which consists of large variations in pose, age, illumination, ethnicity and profession. Cao et al. first trained a SENet with MS-celeb-1M dataset and then fine-tuned the model with VGGface2 , and achieved the SOTA performance on the IJB-A and IJB-B . ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_30",
"text": " Light-weight networks. Using deeper neural network with hundreds of layers and millions of parameters to achieve higher accuracy comes at cost. Powerful GPUs with larger memory size are needed, which makes the applications on many mobiles and embedded devices impractical. To address this problem, light-weight networks are proposed. Light CNN (85, 86) proposed a max-feature-map (MFM) activation function that introduces the concept of maxout in the fully connected layer to CNN. The MFM obtains a compact representation and reduces the computational cost. Sun et al. proposed to sparsify deep networks iteratively from the previously learned denser models based on a weight selection criterion. MobiFace adopted fast downsampling and bottleneck residual block with the expansion layers and achieved high performance with 99.7% on LFW database. Although some other light-weight CNNs, such as SqueezeNet, MobileNet, ShuffleNet and Xception (126, 127, 128, 129), are still not widely used in FR, they deserve more attention. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_31",
"text": " Adaptive-architecture networks. Considering that designing architectures manually by human experts are time-consuming and error-prone processes, there is growing interest in adaptive-architecture networks which can find well-performing architectures, e.g. the type of operation every layer executes (pooling, convolution, etc) and hyper-parameters associated with the operation (number of filters, kernel size and strides for a convolutional layer, etc), according to the specific requirements of training and testing data. Currently, neural architecture search (NAS) is one of the promising methodologies, which has outperformed manually designed architectures on some tasks such as image classification or semantic segmentation . Zhu et al. integrated NAS technology into face recognition. They used reinforcement learning algorithm (policy gradient) to guide the controller network to train the optimal child architecture. Besides NAS, there are some other explorations to learn optimal architectures adaptively. For example, conditional convolutional neural network (c-CNN) dynamically activated sets of kernels according to modalities of samples; Han et al. proposed a novel contrastive convolution consisted of a trunk CNN and a kernel generator, which is beneficial owing to its dynamistic generation of contrastive kernels based on the pair of faces being compared. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_32",
"text": " Joint alignment-recognition networks. Recently, an end-to-end system (91, 92, 93, 94) was proposed to jointly train FR with several modules (face detection, alignment, and so forth) together. Compared to the existing methods in which each module is generally optimized separately according to different objectives, this end-to-end system optimizes each module according to the recognition objective, leading to more adequate and robust inputs for the recognition model. For example, inspired by spatial transformer , Hayat et al. proposed a CNN-based data-driven approach that learns to simultaneously register and represent faces (Fig. 10), while Wu et al. designed a novel recursive spatial transformer (ReST) module for CNN allowing face alignment and recognition to be jointly optimized. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_33",
"text": " Multi-input networks. In “one-to-many augmentation”, multiple images with variety are generated from one image in order to augment training data. Taken these multiple images as input, multiple networks are also assembled together to extract and combine features of different type of inputs, which can outperform an individual network. In (58, 59, 60, 99, 34, 21, 35), assembled networks are built after different face patches are cropped, and then different types of patches are fed into different sub-networks for representation extraction. By combining the results of sub-networks, the performance can be improved. Other papers (96, 95, 98) used assembled networks to recognize images with different poses. For example, Masi et al. adjusted the pose to frontal (0∘superscript00^{\\circ}), half-profile (40∘superscript4040^{\\circ}) and full-profile views (75∘superscript7575^{\\circ}) and then addressed pose variation by assembled pose networks. A multi-view deep network (MvDN) consists of view-specific subnetworks and common subnetworks; the former removes view-specific variations, and the latter obtains common representations. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_34",
"text": " Multi-task networks. FR is intertwined with various factors, such as pose, illumination, and age. To solve this problem, multitask learning is introduced to transfer knowledge from other relevant tasks and to disentangle nuisance factors. In multi-task networks, identity classification is the main task and the side tasks are pose, illumination, and expression estimations, among others. The lower layers are shared among all the tasks, and the higher layers are disentangled into different sub-networks to generate the task-specific outputs. In , the task-specific sub-networks are branched out to learn face detection, face alignment, pose estimation, gender recognition, smile detection, age estimation and FR. Yin et al. proposed to automatically assign the dynamic loss weights for each side task. Peng et al. used a feature reconstruction metric learning to disentangle a CNN into sub-networks for jointly learning the identity and non-identity features as shown in Fig. 11. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_35",
"text": " During testing, the cosine distance and L2 distance are generally employed to measure the similarity between the deep features x1subscript𝑥1x_{1} and x2subscript𝑥2x_{2}; then, threshold comparison and the nearest neighbor (NN) classifier are used to make decision for verification and identification. In addition to these common methods, there are some other explorations. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_36",
"text": " Metric learning, which aims to find a new metric to make two classes more separable, can also be used for face matching based on extracted deep features. The JB model is a well-known metric learning method (35, 21, 36, 34, 120), and Hu et al. proved that it can improve the performance greatly. In the JB model, a face feature x𝑥x is modeled as x=μ+ε𝑥𝜇𝜀x=\\mu+\\varepsilon, where μ𝜇\\mu and ε𝜀\\varepsilon are identity and intra-personal variations, respectively. The similarity score r(x1,x2)𝑟subscript𝑥1subscript𝑥2r(x_{1},x_{2}) can be represented as follows: r(x1,x2)=logP(x1,x2|HI)P(x1,x2|HE)𝑟subscript𝑥1subscript𝑥2𝑙𝑜𝑔𝑃subscript𝑥1conditionalsubscript𝑥2subscript𝐻𝐼𝑃subscript𝑥1conditionalsubscript𝑥2subscript𝐻𝐸r(x_{1},x_{2})=log\\frac{P\\left(x_{1},x_{2}|H_{I}\\right)}{P\\left(x_{1},x_{2}|H_{E}\\right)} (7) where P(x1,x2|HI)𝑃subscript𝑥1conditionalsubscript𝑥2subscript𝐻𝐼P(x_{1},x_{2}|H_{I}) is the probability that two faces belong to the same identity and P(x1,x2|HE)𝑃subscript𝑥1conditionalsubscript𝑥2subscript𝐻𝐸P(x_{1},x_{2}|H_{E}) is the probability that two faces belong to different identities. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_37",
"text": " After cosine distance was computed, Cheng et al. proposed a heuristic voting strategy at the similarity score level to combine the results of multiple CNN models and won first place in Challenge 2 of MS-celeb-1M 2017. Yang et al. extracted the local adaptive convolution features from the local regions of the face image and used the extended SRC for FR with a single sample per person. Guo et al. combined deep features and the SVM classifier to perform recognition. Wang et al. first used product quantization (PQ) to directly retrieve the top-k most similar faces and re-ranked these faces by combining similarities from deep features and the COTS matcher . In addition, Softmax can be also used in face matching when the identities of training set and test set overlap. For example, in Challenge 2 of MS-celeb-1M, Ding et al. trained a 21,000-class softmax classifier to directly recognize faces of one-shot classes and normal classes after augmenting feature by a conditional GAN; Guo et al. trained the softmax classifier combined with underrepresented-classes promotion (UP) loss term to enhance the performance on one-shot classes. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_38",
"text": " When the distributions of training data and testing data are the same, the face matching methods mentioned above are effective. However, there is always a distribution change or domain shift between two data domains that can degrade the performance on test data. Transfer learning (144, 145) has recently been introduced into deep FR to address the problem of domain shift. It learns transferable features using a labeled source domain (training data) and an unlabeled target domain (testing data) such that domain discrepancy is reduced and models trained on source domain will also perform well on target domain. Sometimes, this technology is applied to face matching. For example, Crosswhite et al. and Xiong et al. adopted template adaptation to the set of media in a template by combining CNN features with template-specific linear SVMs. But most of the time, it is not enough to do transfer learning only at face matching stage. Transfer learning should be embedded in deep models to learn more transferable representations. Kan et al. proposed a bi-shifting autoencoder network (BAE) for domain adaptation across view angle, ethnicity, and imaging sensor; while Luo et al. utilized the multi-kernels maximum mean discrepancy (MMD) to reduce domain discrepancies. Sohn et al. used adversarial learning to transfer knowledge from still image FR to video FR. Moreover, fine-tuning the CNN parameters from a prelearned model using a target training dataset is a particular type of transfer learning, and is commonly employed by numerous methods (151, 152, 103). ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_39",
"text": " We present the development of face processing methods in chronological order in Fig. 12. As we can see from the figure, most papers attempted to perform face processing by autoencoder model in 2014 and 2015; while 3D model played an important role in 2016. GAN has drawn substantial attention from the deep learning and computer vision community since it was first proposed by Goodfellow et al. It can be used in different fields and was also introduced into face processing in 2017. GAN can be used to perform “one-to-many augmentation” and “many-to-one normalization”, and it broke the limit that face synthesis should be done under supervised way. Although GAN has not been widely used in face processing for training and recognition, it has great latent capacity for preprocessing, for example, Dual-Agent GANs (DA-GAN) won the 1st places on verification and identification tracks in the NIST IJB-A 2017 FR competitions. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_40",
"text": " Collecting a large database is extremely expensive and time consuming. The methods of “one-to-many augmentation” can mitigate the challenges of data collection, and they can be used to augment not only training data but also the gallery of test data. we categorized them into four classes: data augmentation, 3D model, autoencoder model and GAN model. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_41",
"text": " Data augmentation. Common data augmentation methods consist of photometric transformations (75, 22) and geometric transformations, such as oversampling (multiple patches obtained by cropping at different scales) , mirroring , and rotating the images. Recently, data augmentation has been widely used in deep FR algorithms (58, 59, 60, 35, 21, 36, 61, 62). for example, Sun et al. cropped 400 face patches varying in positions, scales, and color channels and mirrored the images. Liu et al. generated seven overlapped image patches centered at different landmarks on the face region and trained them with seven CNNs with the same structure. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_42",
"text": " 3D model. 3D face reconstruction is also a way to enrich the diversity of training data. They utilize 3D structure information to model the transformation between poses. 3D models first use 3D face data to obtain morphable displacement fields and then apply them to obtain 2D face data in different pose angles. There is a large number of papers about this domain, but we only focus on the 3D face reconstruction using deep methods or used for deep FR. In , Masi et al. generated face images with new intra-class facial appearance variations, including pose, shape and expression, and then trained a 19-layer VGGNet with both real and augmented data. Masi et al. used generic 3D faces and rendered fixed views to reduce much of the computational effort. Richardson et al. employed an iterative 3D CNN by using a secondary input channel to represent the previous network’s output as an image for reconstructing a 3D face as shown in Fig. 13. Dou et al. used a multi-task CNN to divide 3D face reconstruction into neutral 3D reconstruction and expressive 3D reconstruction. Tran et al. directly regressed 3D morphable face model (3DMM) parameters from an input photo by a very deep CNN architecture. An et al. synthesized face images with various poses and expressions using the 3DMM method, then reduced the gap between synthesized data and real data with the help of MMD. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_43",
"text": " Autoencoder model. Rather than reconstructing 3D models from a 2D image and projecting it back into 2D images of different poses, autoencoder models can generate 2D target images directly. Taken a face image and a pose code encoding a target pose as input, an encoder first learns pose-invariant face representation, and then a decoder generates a face image with the same identity viewed at the target pose by using the pose-invariant representation and the pose code. For example, given the target pose codes, multi-view perceptron (MVP) trained some deterministic hidden neurons to learn pose-invariant face representations, and simultaneously trained some random hidden neurons to capture pose features, then a decoder generated the target images by combining pose-invariant representations with pose features. As shown in Fig. 14, Yim et al. and Qian et al. introduced an auxiliary CNN to generate better images viewed at the target poses. First, an autoencoder generated the desired pose image, then the auxiliary CNN reconstructed the original input image back from the generated target image, which guarantees that the generated image is identity-preserving. In , two groups of units are embedded between encoder and decoder. The identity units remain unchanged and the rotation of images is achieved by taking actions to pose units at each time step. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_44",
"text": " GAN model. In GAN models, a generator aims to fool a discriminator through generating images that resemble the real images, while the discriminator aims to discriminate the generated samples from the real ones. By this minimax game between generator and discriminator, GAN can successfully generate photo-realistic images with different poses. After using a 3D model to generate profile face images, DA-GAN refined the images by a GAN, which combines prior knowledge of the data distribution and knowledge of faces (pose and identity perception loss). CVAE-GAN combined a variational auto-encoder with a GAN for augmenting data, and took advantages of both statistic and pairwise feature matching to make the training process converge faster and more stably. In addition to synthesizing diverse faces from noise, some papers also explore to disentangle the identity and variation, and synthesize new faces by exchanging identity and variation from different people. In CG-GAN , a generator directly resolves each representation of input image into a variation code and an identity code and regroups these codes for cross-generating, simultaneously, a discriminator ensures the reality of generated images. Bao et al. extracted identity representation of one input image and attribute representation of any other input face image, then synthesized new faces by recombining these representations. This work shows superior performance in generating realistic and identity preserving face images, even for identities outside the training dataset. Unlike previous methods that treat classifier as a spectator, FaceID-GAN proposed a three-player GAN where the classifier cooperates together with the discriminator to compete with the generator from two different aspects, i.e. facial identity and image quality respectively. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_45",
"text": " In contrast to “one-to-many augmentation”, the methods of “many-to-one normalization” produce frontal faces and reduce appearance variability of test data to make faces align and compare easily. It can be categorized as autoencoder model, CNN model and GAN model. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_46",
"text": " Autoencoder model. Autoencoder can also be applied to “many-to-one normalization”. Different from the autoencoder model in “one-to-many augmentation” which generates the desired pose images with the help of pose codes, autoencoder model here learns pose-invariant face representation by an encoder and directly normalizes faces by a decoder without pose codes. Zhu et al. (66, 67) selected canonical-view images according to the face images’ symmetry and sharpness and then adopted an autoencoder to recover the frontal view images by minimizing the reconstruction loss error. The proposed stacked progressive autoencoders (SPAE) progressively map the nonfrontal face to the frontal face through a stack of several autoencoders. Each shallow autoencoders of SPAE is designed to convert the input face images at large poses to a virtual view at a smaller pose, so the pose variations are narrowed down gradually layer by layer along the pose manifold. Zhang et al. built a sparse many-to-one encoder to enhance the discriminant of the pose free feature by using multiple random faces as the target values for multiple encoders. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_47",
"text": " CNN model. CNN models usually directly learn the 2D mappings between non-frontal face images and frontal images, and utilize these mapping to normalize images in pixel space. The pixels in normalized images are either directly the pixels or the combinations of the pixels in non-frontal images. In LDF-Net , the displacement field network learns the shifting relationship of two pixels, and the translation layer transforms the input non-frontal face image into a frontal one with this displacement field. In GridFace shown in Fig. 15, first, the rectification network normalizes the images by warping pixels from the original image to the canonical one according to the computed homography matrix, then the normalized output is regularized by an implicit canonical view face prior, finally, with the normalized faces as input, the recognition network learns discriminative face representation via metric learning. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_48",
"text": " GAN model. Huang et al. proposed a two-pathway generative adversarial network (TP-GAN) that contains four landmark-located patch networks and a global encoder-decoder network. Through combining adversarial loss, symmetry loss and identity-preserving loss, TP-GAN generates a frontal view and simultaneously preserves global structures and local details as shown in Fig. 16. In a disentangled representation learning generative adversarial network (DR-GAN) , the generator serves as a face rotator, in which an encoder produces an identity representation, and a decoder synthesizes a face at the specified pose using this representation and a pose code. And the discriminator is trained to not only distinguish real vs. synthetic images, but also predict the identity and pose of a face. Yin et al. incorporated 3DMM into the GAN structure to provide shape and appearance priors to guide the generator to frontalization. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_49",
"text": " In the past three decades, many face databases have been constructed with a clear tendency from small-scale to large-scale, from single-source to diverse-sources, and from lab-controlled to real-world unconstrained condition, as shown in Fig. 17. As the performance of some simple databases become saturated, e.g. LFW , more and more complex databases were continually developed to facilitate the FR research. It can be said without exaggeration that the development process of the face databases largely leads the direction of FR research. In this section, we review the development of major training and testing academic databases for the deep FR. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_50",
"text": " The prerequisite of effective deep FR is a sufficiently large training dataset. Zhou et al. suggested that large amounts of data with deep learning improve the performance of FR. The results of Megaface Challenge also revealed that premier deep FR methods were typically trained on data larger than 0.5M images and 20K people. The early works of deep FR were usually trained on private training datasets. Facebook’s Deepface model was trained on 4M images of 4K people; Google’s FaceNet was trained on 200M images of 3M people; DeepID serial models (34, 35, 21, 36) were trained on 0.2M images of 10K people. Although they reported ground-breaking performance at this stage, researchers cannot accurately reproduce or compare their models without public training datasets. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_51",
"text": " To address this issue, CASIA-Webface provided the first widely-used public training dataset for the deep model training purpose, which consists of 0.5M images of 10K celebrities collected from the web. Given its moderate size and easy usage, it has become a great resource for fair comparisons for academic deep models. However, its relatively small data and ID size may not be sufficient to reflect the power of many advanced deep learning methods. Currently, there have been more databases providing public available large-scale training dataset (Table VI), especially three databases with over 1M images, namely MS-Celeb-1M , VGGface2 , and Megaface (44, 164), and we summary some interesting findings about these training sets, as shown in Fig. 18. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_52",
"text": " Depth v.s. breadth. These large training sets are expanded from depth or breadth. VGGface2 provides a large-scale training dataset of depth, which have limited number of subjects but many images for each subjects. The depth of dataset enforces the trained model to address a wide range intra-class variations, such as lighting, age, and pose. In contrast, MS-Celeb-1M and Mageface (Challenge 2) offers large-scale training datasets of breadth, which contains many subject but limited images for each subjects. The breadth of dataset ensures the trained model to cover the sufficiently variable appearance of various people. Cao et al. conducted a systematic studies on model training using VGGface2 and MS-Celeb-1M, and found an optimal model by first training on MS-Celeb-1M (breadth) and then fine-tuning on VGGface2 (depth). ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_53",
"text": " Long tail distribution. The utilization of long tail distribution is different among datasets. For example, in Challenge 2 of MS-Celeb-1M, the novel set specially uses the tailed data to study low-shot learning; central part of the long tail distribution is used by the Challenge 1 of MS-Celeb-1M and images’ number is approximately limited to 100 for each celebrity; VGGface and VGGface2 only use the head part to construct deep databases; Megaface utilizes the whole distribution to contain as many images as possible, the minimal number of images is 3 per person and the maximum is 2469. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_54",
"text": " Data engineering. Several popular benchmarks, such as LFW unrestricted protocol, Megaface Challenge 1, MS-Celeb-1M Challenge 1&2, explicitly encourage researchers to collect and clean a large-scale data set for enhancing the capability of deep neural network. Although data engineering is a valuable problem to computer vision researchers, this protocol is more incline to the industry participants. As evidence, the leaderboards of these experiments are mostly occupied by the companies holding invincible hardwares and data scales. This phenomenon may not be beneficial for developments of new models in academic community. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_55",
"text": " Data noise. Owing to data source and collecting strategies, existing large-scale datasets invariably contain label noises. Wang et al. profiled the noise distribution in existing datasets in Fig. 19 and showed that the noise percentage increases dramatically along the scale of data. Moreover, they found that noise is more lethal on a 10,000-class problem of FR than on a 10-class problem of object classification and that label flip noise severely deteriorates the performance of a model, especially the model using A-softmax . Therefore, building a sufficiently large and clean dataset for academic research is very meaningful. Deng et al. found there are serious label noise in MS-Celeb-1M , and they cleaned the noise of MS-Celeb-1M, and made the refined dataset public available. Microsoft and Deepglint jointly released the largest public data set with cleaned labels, which includes 4M images cleaned from MS-Celeb-1M dataset and 2.8M aligned images of 100K Asian celebrities. Moreover, Zhan et al. shifted the focus from cleaning the datasets to leveraging more unlabeled data. Through automatically assigning pseudo labels to unlabeled data with the help of relational graphs, they obtained competitive or even better results over the fully-supervised counterpart. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_56",
"text": " Data bias. Large-scale training datasets, such as CASIA-WebFace , VGGFace2 and MS-Celeb-1M , are typically constructed by scraping websites like Google Images, and consist of celebrities on formal occasions: smiling, make-up, young, and beautiful. They are largely different from databases captured in the daily life (e.g. Megaface). The biases can be attributed to many exogenous factors in data collection, such as cameras, lightings, preferences over certain types of backgrounds, or annotator tendencies. Dataset biases adversely affect cross-dataset generalization; that is, the performance of the model trained on one dataset drops significantly when applied to another one. One persuasive evidence is presented by P.J. Phillips’ study which conducted a cross benchmark assessment of VGGFace model for face recognition. The VGGFace model achieves 98.95% on LFW and 97.30% on YTF , but only obtains 26%, 52% and 85% on Ugly, Bad and Good partition of GBU database . ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_57",
"text": " Demographic bias (e.g., race/ethnicity, gender, age) in datasets is a universal but urgent issue to be solved in data bias field. In existing training and testing datasets, the male, White, and middle-aged cohorts always appear more frequently, as shown in Table VII, which inevitably causes deep learning models to replicate and even amplify these biases resulting in significantly different accuracies when deep models are applied to different demographic groups. Some researches (145, 171, 172) showed that the female, Black, and younger cohorts are usually more difficult to recognize in FR systems trained with commonly-used datasets. For example, Wang et al. proposed a Racial Faces in-the-Wild (RFW) database and proved that existing commercial APIs and the SOTA algorithms indeed work unequally for different races and the maximum difference in error rate between the best and worst groups is 12%, as shown in Table VIII. Hupont et al. showed that SphereFace has a TAR of 0.87 for White males which drops to 0.28 for Asian females, at a FAR of 1e−41𝑒41e-4. Such bias can result in mistreatment of certain demographic groups, by either exposing them to a higher risk of fraud, or by making access to services more difficult. Therefore, addressing data bias and enhancing fairness of FR systems in real life are urgent and necessary tasks. Collecting balanced data to train a fair model or designing some debiasing algorithms are effective way. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_58",
"text": " In terms of training protocol, FR can be categorized into subject-dependent and subject-independent settings, as illustrated in Fig. 20. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_59",
"text": " Subject-dependent protocol. For subject-dependent protocol, all testing identities are predefined in training set, it is natural to classify testing face images to the given identities. Therefore, subject-dependent FR can be well addressed as a classification problem, where features are expected to be separable. The protocol is mostly adopted by the early-stage (before 2010) FR studies on FERET , AR , and is suitable only for some small-scale applications. The Challenge 2 of MS-Celeb-1M is the only large-scale database using subject-dependent training protocol. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_60",
"text": " Subject-independent protocol. For subject-independent protocol, the testing identities are usually disjoint from the training set, which makes FR more challenging yet close to practice. Because it is impossible to classify faces to known identities in training set, generalized representation is essential. Due to the fact that human faces exhibit similar intra-subject variations, deep models can display transcendental generalization ability when training with a sufficiently large set of generic subjects, where the key is to learn discriminative large-margin deep features. This generalization ability makes subject-independent FR possible. Almost all major face-recognition benchmarks, such as LFW , PaSC , IJB-A/B/C (41, 42, 43) and Megaface (44, 164), require the tested models to be trained under subject-independent protocol. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_61",
"text": " In order to evaluate whether our deep models can solve the different problems of FR in real life, many testing datasets are designed to evaluate the models in different tasks, i.e. face verification, close-set face identification and open-set face identification. In either task, a set of known subjects is initially enrolled in the system (the gallery), and during testing, a new subject (the probe) is presented. Face verification computes one-to-one similarity between the gallery and probe to determine whether the two images are of the same subject, whereas face identification computes one-to-many similarity to determine the specific identity of a probe face. When the probe appears in the gallery identities, this is referred to as closed-set identification; when the probes include those who are not in the gallery, this is open-set identification. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_62",
"text": " Face verification is relevant to access control systems, re-identification, and application independent evaluations of FR algorithms. It is classically measured using the receiver operating characteristic (ROC) and estimated mean accuracy (Acc). At a given threshold (the independent variable), ROC analysis measures the true accept rate (TAR), which is the fraction of genuine comparisons that correctly exceed the threshold, and the false accept rate (FAR), which is the fraction of impostor comparisons that incorrectly exceed the threshold. And Acc is a simplified metric introduced by LFW , which represents the percentage of correct classifications. With the development of deep FR, more accurate recognitions are required. Customers concern more about the TAR when FAR is kept in a very low rate in most security certification scenario. PaSC reports TAR at a FAR of 10−2superscript10210^{-2}; IJB-A evaluates TAR at a FAR of 10−3superscript10310^{-3}; Megaface (44, 164) focuses on TAR@10−6superscript10610^{-6}FAR; especially, in MS-celeb-1M challenge 3 , TAR@10−9superscript10910^{-9}FAR is reported. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_63",
"text": " Close-set face identification is relevant to user driven searches (e.g., forensic identification), rank-N and cumulative match characteristic (CMC) is commonly used metrics in this scenario. Rank-N is based on what percentage of probe searches return the probe’s gallery mate within the top k𝑘k rank-ordered results. The CMC curve reports the percentage of probes identified within a given rank (the independent variable). IJB-A/B/C (41, 42, 43) concern on the rank-1 and rank-5 recognition rate. The MegaFace challenge (44, 164) systematically evaluates rank-1 recognition rate function of increasing number of gallery distractors (going from 10 to 1 Million), the results of the SOTA evaluated on MegaFace challenge are listed in Table IX. Rather than rank-N and CMC, MS-Celeb-1M further applies a precision-coverage curve to measure identification performance under a variable threshold t𝑡t. The probe is rejected when its confidence score is lower than t𝑡t. The algorithms are compared in term of what fraction of passed probes, i.e. coverage, with a high recognition precision, e.g. 95% or 99%, the results of the SOTA evaluated on MS-Celeb-1M challenge are listed in Table X. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_64",
"text": " Open-set face identification is relevant to high throughput face search systems (e.g., de-duplication, watch list identification), where the recognition system should reject unknown/unseen subjects (probes who do not present in gallery) at test time. At present, there are very few databases covering the task of open-set FR. IJB-A/B/C (41, 42, 43) benchmarks introduce a decision error tradeoff (DET) curve to characterize the the false negative identification rate (FNIR) as function of the false positive identification rate (FPIR). FPIR measures what fraction of comparisons between probe templates and non-mate gallery templates result in a match score exceeding T𝑇T. At the same time, FNIR measures what fraction of probe searches will fail to match a mated gallery template above a score of T𝑇T. The algorithms are compared in term of the FNIR at a low FPIR, e.g. 1% or 10%, the results of the SOTA evaluated on IJB-A dataset as listed in Table XI. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_65",
"text": " Public available training databases are mostly collected from the photos of celebrities due to privacy issue, it is far from images captured in the daily life with diverse scenes. In order to study different specific scenarios, more difficult and realistic datasets are constructed accordingly, as shown in Table XII. According to their characteristics, we divide these scenes into four categories: cross-factor FR, heterogenous FR, multiple (or single) media FR and FR in industry (Fig. 21). ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_66",
"text": " • Cross-factor FR. Due to the complex nonlinear facial appearance, some variations will be caused by people themselves, such as cross-pose, cross-age, make-up, and disguise. For example, CALFW , MORPH , CACD and FG-NET are commonly used datasets with different age range; CFP only focuses on frontal and profile face, CPLFW is extended from LFW and contains different poses. Disguised faces in the wild (DFW) evaluates face recognition across disguise. • Heterogenous FR. It refers to the problem of matching faces across different visual domains. The domain gap is mainly caused by sensory devices and cameras settings, e.g. visual light vs. near-infrared and photo vs. sketch. For example, CUFSF and CUFS are commonly used photo-sketch datasets and CUFSF dataset is harder due to lighting variation and shape exaggeration. • Multiple (or single) media FR. Ideally, in FR, many images of each subject are provided in training datasets and image-to-image recognitions are performed when testing. But the situation will be different in reality. Sometimes, the number of images per person in training set could be very small, such as MS-Celeb-1M challenge 2 . This challenge is often called low- shot or few-shot FR. Moreover, each subject face in test set may be enrolled with a set of images and videos and set-to-set recognition should be performed, such as IJB-A and PaSC . • FR in industry. Although deep FR has achieved beyond human performance on some standard benchmarks, but some other factors should be given more attention rather than accuracy when deep FR is adopted in industry, e.g. anti-attack (CASIA-FASD ) and 3D FR (Bosphorus , BU-3DFE and FRGCv2 ). Compared to publicly available 2D face databases, 3D scans are hard to acquire, and the number of scans and subjects in public 3D face databases is still limited, which hinders the development of 3D deep FR. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_67",
"text": " Despite the high accuracy in the LFW and Megaface (44, 164) benchmarks, the performance of FR models still hardly meets the requirements in real-world application. A conjecture in industry is made that results of generic deep models can be improved simply by collecting big datasets of the target scene. However, this holds only to a certain degree. More and more concerns on privacy may make the collection and human-annotation of face data become illegal in the future. Therefore, significant efforts have been paid to design excellent algorithms to address the specific problems with limited data in these realistic scenes. In this section, we present several special algorithms of FR. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_68",
"text": " As shows that many existing algorithms suffer a decrease of over 10% from frontal-frontal to frontal-profile verification, cross-pose FR is still an extremely challenging scene. In addition to the aforementioned methods, including “one-to-many augmentation”, “many-to-one normalization” and assembled networks (Section 4 and 3.2.2), there are some other algorithms designed for cross-pose FR. Considering the extra burden of above methods, Cao et al. attempted to perform frontalization in the deep feature space rather than the image space. A deep residual equivariant mapping (DREAM) block dynamically added residuals to an input representation to transform a profile face to a frontal image. Chen et al. proposed to combine feature extraction with multi-view subspace learning to simultaneously make features be more pose-robust and discriminative. Pose Invariant Model (PIM) jointly performed face frontalization and learned pose invariant representations end-to-end to allow them to mutually boost each other, and further introduced unsupervised cross-domain adversarial training and a learning to learn strategy to provide high-fidelity frontal reference face images. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_69",
"text": " Cross-age FR is extremely challenging due to the changes in facial appearance by the aging process over time. One direct approach is to synthesize the desired image with target age such that the recognition can be performed in the same age group. A generative probabilistic model was used by to model the facial aging process at each short-term stage. The identity-preserved conditional generative adversarial networks (IPCGANs) framework utilized a conditional-GAN to generate a face in which an identity-preserved module preserved the identity information and an age classifier forced the generated face with the target age. Antipov et al. proposed to age faces by GAN, but the synthetic faces cannot be directly used for face verification due to its imperfect preservation of identities. Then, they used a local manifold adaptation (LMA) approach to solve the problem of . In , high-level age-specific features conveyed by the synthesized face are estimated by a pyramidal adversarial discriminator at multiple scales to generate more lifelike facial details. An alternative to address the cross-age problem is to decompose aging and identity components separately and extract age-invariant representations. Wen et al. developed a latent identity analysis (LIA) layer to separate these two components, as shown in Fig. 22. In , age-invariant features were obtained by subtracting age-specific factors from the representations with the help of the age estimation task. In , face features are decomposed in the spherical coordinate system, in which the identity-related components are represented with angular coordinates and the age-related information is encoded with radial coordinate. Additionally, there are other methods designed for cross-age FR. For example, Bianco ett al. and El et al. fine-tuned the CNN to transfer knowledge across age. Wang et al. proposed a siamese deep network to perform multi-task learning of FR and age estimation. Li et al. integrated feature extraction and metric learning via a deep CNN. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_70",
"text": " Makeup is widely used by the public today, but it also brings challenges for FR due to significant facial appearance changes. The research on matching makeup and nonmakeup face images is receiving increasing attention. Li et al. generated nonmakeup images from makeup ones by a bi-level adversarial network (BLAN) and then used the synthesized nonmakeup images for verification as shown in Fig. 23. Sun et al. pretrained a triplet network on videos and fine-tuned it on a small makeup datasets. Specially, facial disguise (214, 228, 229) is a challenging research topic in makeup face recognition. By using disguise accessories such as wigs, beard, hats, mustache, and heavy makeup, disguise introduces two variations: (i) when a person wants to obfuscate his/her own identity, and (ii) another individual impersonates someone else’s identity. Obfuscation increases intra-class variations whereas impersonation reduces the inter-class dissimilarity, thereby affecting face recognition/verification task. To address this issue, a variety of methods are proposed. Zhang et al. first trained two DCNNs for generic face recognition and then used Principal Components Analysis (PCA) to find the transformation matrix for disguised face recognition adaptation. Kohli et al. finetuned models using disguised faces. Smirnov et al. proposed a hard example mining method benefitted from class-wise (Doppelganger Mining ) and example-wise mining to learn useful deep embeddings for disguised face recognition. Suri et al. learned the representations of images in terms of colors, shapes, and textures (COST) using an unsupervised dictionary learning method, and utilized the combination of COST features and CNN features to perform recognition. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_71",
"text": " Due to the excellent performance of the near-infrared spectrum (NIS) images under low-light scenarios, NIS images are widely applied in surveillance systems. Because most enrolled databases consist of visible light (VIS) spectrum images, how to recognize a NIR face from a gallery of VIS images has been a hot topic. Saxena et al. and Liu et al. transferred the VIS deep networks to the NIR domain by fine-tuning. Lezama et al. used a VIS CNN to recognize NIR faces by transforming NIR images to VIS faces through cross-spectral hallucination and restoring a low-rank structure for features through low-rank embedding. Reale et al. trained a VISNet (for visible images) and a NIRNet (for near-infrared images), and coupled their output features by creating a siamese network. He et al. (238, 239) divided the high layer of the network into a NIR layer, a VIS layer and a NIR-VIS shared layer, then, a modality-invariant feature can be learned by the NIR-VIS shared layer. Song et al. embedded cross-spectral face hallucination and discriminative feature learning into an end-to-end adversarial network. In , the low-rank relevance and cross-modal ranking were used to alleviate the semantic gap. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_72",
"text": " Although deep networks are robust to low resolution to a great extent, there are still a few studies focused on promoting the performance of low-resolution FR. For example, Zangeneh et al. proposed a CNN with a two-branch architecture (a super-resolution network and a feature extraction network) to map the high- and low-resolution face images into a common space where the intra-person distance is smaller than the inter-person distance. Shen et al. exploited the face semantic information and local structural constraints to better restore the shape and detail of face images. In addition, they optimized the network with perceptual and adversarial losses to produce photo-realistic results. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_73",
"text": " The photo-sketch FR may help law enforcement to quickly identify suspects. The commonly used methods can be categorized as two classes. One is to utilize transfer learning to directly match photos to sketches. Deep networks are first trained using a large face database of photos and are then fine-tuned using small sketch database (243, 244). The other is to use the image-to-image translation, where the photo can be transformed to a sketch or the sketch to a photo; then, FR can be performed in one domain. Zhang et al. developed a fully convolutional network with generative loss and a discriminative regularizer to transform photos to sketches. Zhang et al. utilized a branched fully convolutional neural network (BFCN) to generate a structure-preserved sketch and a texture-preserved sketch, and then they fused them together via a probabilistic method. Recently, GANs have achieved impressive results in image generation. Yi et al. , Kim et al. and Zhu et al. used two generators, GAsubscript𝐺𝐴G_{A} and GBsubscript𝐺𝐵G_{B}, to generate sketches from photos and photos from sketches, respectively (Fig. 24). Based on , Wang et al. proposed a multi-adversarial network to avoid artifacts by leveraging the implicit presence of feature maps of different resolutions in the generator subnetwork. Similar to photo-sketch FR, photo-caricature FR is one kind of heterogenous FR scenes which is challenging and important to understanding of face perception. Huo et al. built a large dataset of caricatures and photos, and provided several evaluation protocols and their baseline performances for comparison. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_74",
"text": " For many practical applications, such as surveillance and security, the FR system should recognize persons with a very limited number of training samples or even with only one sample. The methods of low-shot learning can be categorized as 1) synthesizing training data and 2) learning more powerful features. Hong et al. generated images in various poses using a 3D face model and adopted deep domain adaptation to handle other variations, such as blur, occlusion, and expression (Fig. 25). Choe et al. used data augmentation methods and a GAN for pose transition and attribute boosting to increase the size of the training dataset. Wu et al. proposed a framework with hybrid classifiers using a CNN and a nearest neighbor (NN) model. Guo et al. made the norms of the weight vectors of the one-shot classes and the normal classes aligned to address the data imbalance problem. Cheng et al. proposed an enforced softmax that contains optimal dropout, selective attenuation, L2 normalization and model-level optimization. Yin et al. augmented feature space of low-shot classes by transferring the principal components from regular to low-shot classes to encourage the variance of low-shot classes to mimic that of regular classes. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_75",
"text": " Different from traditional image-to-image recognition, set-to-set recognition takes a set (heterogeneous contents containing both images and videos) as the smallest unit of representation. This kind of setting does reflect the real-world biometric scenarios, thereby attracting a lot of attention. After learning face representations of media in each set, two strategies are generally adopted to perform set-to-set matching. One is to use these representations to perform pair-wise similarity comparison of two sets and aggregate the results into a single and final score by max score pooling , average score pooling and its variations (253, 254). The other strategy is feature pooling (96, 103, 81) which first aggregates face representations into a single representation for each set and then performs a comparison between two sets. In addition to the commonly used strategies, there are also some novel methods proposed for set/template-based FR. For example, Hayat et al. proposed a deep heterogeneous feature fusion network to exploit the features’ complementary information generated by different CNNs. Liu et al. introduced the actor-critic reinforcement learning for set-based FR. They casted the inner-set dependency modeling to a Markov decision process in the latent space, and trained a dependency-aware attention control agent to make attention control for each image in each step. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_76",
"text": " There are two key issues in video FR: one is to integrate the information across different frames together to build a representation of the video face, and the other is to handle video frames with severe blur, pose variations, and occlusions. For frame aggregation, Yang et al. proposed a neural aggregation network (NAN) in which the aggregation module, consisting of two attention blocks driven by a memory, produces a 128-dimensional vector representation (Fig. 26). Rao et al. aggregated raw video frames directly by combining the idea of metric learning and adversarial learning. For dealing with bad frames, Rao et al. discarded the bad frames by treating this operation as a Markov decision process and trained the attention model through a deep reinforcement learning framework. Ding et al. artificially blurred clear images for training to learn blur-robust face representations. Parchami et al. used a CNN to reconstruct a lower-quality video into a high-quality face. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_77",
"text": " 3D FR has inherent advantages over 2D methods, but 3D deep FR is not well developed due to the lack of large annotated 3D data. To enlarge 3D training datasets, most works use the methods of “one-to-many augmentation” to synthesize 3D faces. However, the effective methods for extracting deep features of 3D faces remain to be explored. Kim et al. fine-tuned a 2D CNN with a small amount of 3D scans for 3D FR. Zulqarnain et al. used a three-channel (corresponding to depth, azimuth and elevation angles of the normal vector) image as input and minimized the average prediction log-loss. Zhang et al. first selected 30 feature points from the Candide-3 face model to characterize faces, then conducted the unsupervised pretraining of face depth data, and finally performed the supervised fine-tuning. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_78",
"text": " Partial FR, in which only arbitrary-size face patches are presented, has become an emerging problem with increasing requirements of identification from CCTV cameras and embedded vision systems in mobile devices, robots and smart home facilities. He et al. divided the aligned face image into several multi-scale patches, and the dissimilarity between two partial face images is calculated as the weighted L2 distance between corresponding patches. Dynamic feature matching (DFM) utilized a sliding window of the same size as the probe feature maps to decompose the gallery feature maps into several gallery sub-feature maps, and the similarity-guided constraint imposed on sparse representation classification (SRC) provides an alignment-free matching. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_79",
"text": " With the emergence of mobile phones, tablets and augmented reality, FR has been applied in mobile devices. Due to computational limitations, the recognition tasks in these devices need to be carried out in a light but timely fashion. MobiFace required efficient memory and low cost operators by adopting fast downsampling and bottleneck residual block, and achieves 99.7% on LFW database and 91.3% on Megaface database. Tadmor et al. proposed a multibatch method that first generates signatures for a minibatch of k𝑘k face images and then constructs an unbiased estimate of the full gradient by relying on all k2−ksuperscript𝑘2𝑘k^{2}-k pairs from the minibatch. As mentioned in Section 3.2.1, light-weight deep networks (126, 127, 128, 129) perform excellently in the fundamental tasks of image classification and deserve further attention in FR tasks. Moreover, some well-known compressed networks such as Pruning (264, 265, 266), BinaryNets (267, 268, 269, 270), Mimic Networks (271, 272), also have potential to be introduced into FR. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_80",
"text": " With the success of FR techniques, various types of attacks, such as face spoofing and adversarial perturbations, are becoming large threats. Face spoofing involves presenting a fake face to the biometric sensor using a printed photograph, worn mask, or even an image displayed on another electronic device. In order to defense this type of attack, several methods are proposed (211, 273, 274, 275, 276, 277, 278, 279). Atoum et al. proposed a novel two-stream CNN in which the local features discriminate the spoof patches that are independent of the spatial face areas, and holistic depth maps ensure that the input live sample has a face-like depth. Yang et al. trained a CNN using both a single frame and multiple frames with five scales as input, and using the live/spoof label as the output. Taken the sequence of video frames as input, Xu et al. applied LSTM units on top of CNN to obtain end-to-end features to recognize spoofing faces which leveraged the local and dense property from convolution operation and learned the temporal structure using LSTM units. Li et al. and Patel et al. fine-tuned their networks from a pretrained model by training sets of real and fake images. Jourabloo et al. proposed to inversely decompose a spoof face into the live face and the spoof noise pattern. Adversarial perturbation is the other type of attack which can be defined as the addition of a minimal vector r𝑟r such that with addition of this vector into the input image x𝑥x, i.e. (x+r)𝑥𝑟(x+r), the deep learning models misclassifies the input while people will not. Recently, more and more work has begun to focus on solving this perturbation of FR. Goswami et al. proposed to detect adversarial samples by characterizing abnormal filter response behavior in the hidden layers and increase the network’s robustness by removing the most problematic filters. Goel et al. provided an open source implementation of adversarial detection and mitigation algorithms. Despite of progresses of anti-attack algorithms, attack methods are updated as well and remind us the need to further increase security and robustness in FR systems, for example, Mai et al. proposed a neighborly de-convolutional neural network (NbNet) to reconstruct a fake face using the stolen deep templates. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_81",
"text": " As described in Section 5.1, existing datasets are highly biased in terms of the distribution of demographic cohorts, which may dramatically impact the fairness of deep models. To address this issue, there are some works that seek to introduce fairness into face recognition and mitigate demographic bias, e,g. unbalanced-training , attribute removal (284, 285, 286) and domain adaptation (173, 287, 147). 1) Unbalanced-training methods mitigate the bias via model regularization, taking into consideration of the fairness goal in the overall model objective function. For example, RL-RBN formulated the process of finding the optimal margins for non-Caucasians as a Markov decision process and employed deep Q-learning to learn policies based on large margin loss. 2) Attribute removal methods confound or remove demographic information of faces to learn attribute-invariant representations. For example, Alvi et al. applied a confusion loss to make a classifier fail to distinguish attributes of examples so that multiple spurious variations are removed from the feature representation. SensitiveNets proposed to introduce sensitive information into triplet loss. They minimized the sensitive information, while maintaining distances between positive and negative embeddings. 3) Domain adaptation methods propose to investigate data bias problem from a domain adaptation point of view and attempt to design domain-invariant feature representations to mitigate bias across domains. IMAN simultaneously aligned global distribution to decrease race gap at domain-level, and learned the discriminative target representations at cluster level. Kan directly converted the Caucasian data to non-Caucasian domain in the image space with the help of sparse reconstruction coefficients learnt in the common subspace. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_82",
"text": " In this paper, we provide a comprehensive survey of deep FR from both data and algorithm aspects. For algorithms, mainstream and special network architectures are presented. Meanwhile, we categorize loss functions into Euclidean-distance-based loss, angular/cosine-margin-based loss and variable softmax loss. For data, we summarize some commonly used datasets. Moreover, the methods of face processing are introduced and categorized as “one-to-many augmentation” and “many-to-one normalization”. Finally, the special scenes of deep FR, including video FR, 3D FR and cross-age FR, are briefly introduced. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_83",
"text": " Taking advantage of big annotated data and revolutionary deep learning techniques, deep FR has dramatically improved the SOTA performance and fostered successful real-world applications. With the practical and commercial use of this technology, many ideal assumptions of academic research were broken, and more and more real-world issues are emerging. To the best our knowledge, major technical challenges include the following aspects. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_84",
"text": " • Security issues. Presentation attack , adversarial attack (280, 281, 290), template attack and digital manipulation attack (292, 293) are developing to threaten the security of deep face recognition systems. 1) Presentation attack with 3D silicone mask, which exhibits skin-like appearance and facial motion, challenges current anti-sproofing methods . 2) Although adversarial perturbation detection and mitigation methods are recently proposed , the root cause of adversarial vulnerability is unclear and thus new types of adversarial attacks are still upgraded continuously (295, 296). 3) The stolen deep feature template can be used to recover its facial appearance, and how to generate cancelable template without loss of accuracy is another important issue. 4) Digital manipulation attack, made feasible by GANs, can generate entirely or partially modified photorealistic faces by expression swap, identity swap, attribute manipulation and entire face synthesis, which remains a main challenge for the security of deep FR. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_85",
"text": " • Privacy-preserving face recognition. With the leakage of biological data, privacy concerns are raising nowadays. Facial images can predict not only demographic information such as gender, age, or race, but even the genetic information . Recently, the pioneer works such as Semi-Adversarial Networks (298, 299, 285) have explored to generate a recognizable biometric templates that can hidden some of the private information presented in the facial images. Further research on the principles of visual cryptography, signal mixing and image perturbation to protect users’ privacy on stored face templates are essential for addressing public concern on privacy. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_86",
"text": " • Understanding deep face recognition. Deep face recognition systems are now believed to surpass human performance in most scenarios . There are also some interesting attempts to apply deep models to assist human operators for face verification . Despite this progress, many fundamental questions are still open, such as what is the “identity capacity” of a deep representation ? Why deep neural networks, rather than humans, are easily fooled by adversarial samples? While bigger and bigger training dataset by itself cannot solve this problem, deeper understanding on these questions may help us to build robust applications in real world. Recently, a new benchmark called TALFW has been proposed to explore this issue . ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_87",
"text": " • Remaining challenges defined by non-saturated benchmark datasets. Three current major datasets, namely, MegaFace (44, 164) , MS-Celeb-1M and IJB-A/B/C (41, 42, 43), are corresponding to large-scale FR with a very large number of candidates, low/one-shot FR and large pose-variance FR which will be the focus of research in the future. Although the SOTA algorithms can be over 99.9 percent accurate on LFW and Megaface (44, 164) databases, fundamental challenges such as matching faces cross ages , poses , sensors, or styles still remain. For both datasets and algorithms, it is necessary to measure and address the racial/gender/age biases of deep FR in future research. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_88",
"text": " • Ubiquitous face recognition across applications and scenes. Deep face recognition has been successfully applied on many user-cooperated applications, but the ubiquitous recognition applications in everywhere are still an ambitious goal. In practice, it is difficult to collect and label sufficient samples for innumerable scenes in real world. One promising solution is to first learn a general model and then transfer it to an application-specific scene. While deep domain adaptation has recently been applied to reduce the algorithm bias on different scenes , different races , general solution to transfer face recognition is largely open. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_89",
"text": " • Pursuit of extreme accuracy and efficiency. Many killer-applications, such as watch-list surveillance or financial identity verification, require high matching accuracy at very low alarm rate, e.g. 10−9superscript10910^{-9}. It is still a big challenge even with deep learning on massive training data. Meanwhile, deploying deep face recognition on mobile devices pursues the minimum size of feature representation and compressed deep network. It is of great significance for both industry and academic to explore this extreme face-recognition performance beyond human imagination. It is also exciting to constantly push the performance limits of the algorithm after it has already surpassed human. ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_90",
"text": " • Fusion issues. Face recognition by itself is far from sufficient to solve all biometric and forensic tasks, such as distinguishing identical twins and matching faces before and after surgery . A reliable solution is to consolidate multiple sources of biometric evidence . These sources of information may correspond to different biometric traits (e.g., face + hand ), sensors (e.g., 2D + 3D face cameras), feature extraction and matching techniques, or instances (e.g., a face sequence of various poses). It is beneficial for face biometric and forensic applications to perform information fusion at the data level, feature level, score level, rank level, and decision level . ",
"title": "Deep Face Recognition"
},
{
"id": "1804.06655_all_91",
"text": " This work was partially supported by National Key R&D Program of China (2019YFB1406504) and BUPT Excellent Ph.D. Students Foundation CX2020207. ",
"title": "Deep Face Recognition"
}
] |
If, for a certain model, it was theorized that the penultimate layer is the most important later for generating embeddings, how could discriminative fine-tuning be used to validate or refute that theory?
|
In this work, discriminative fine-tuning was used to fine-tune each layer with a different learning rate [17]. Specifically, the learning rate was decreased going from the last layer to lower layers [45]. The authors found that this improved performance across several datasets [19].
|
[
17,
45,
19
] |
[
{
"id": "1801.06146_all_0",
"text": " Inductive transfer learning has had a large impact on computer vision (CV). Applied CV models (including object detection, classification, and segmentation) are rarely trained from scratch, but instead are fine-tuned from models that have been pretrained on ImageNet, MS-COCO, and other datasets Sharif Razavian et al. (2014); Long et al. (2015a); He et al. (2016); Huang et al. (2017). ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_1",
"text": " Text classification is a category of Natural Language Processing (NLP) tasks with real-world applications such as spam, fraud, and bot detection Jindal and Liu (2007); Ngai et al. (2011); Chu et al. (2012), emergency response Caragea et al. (2011), and commercial document classification, such as for legal discovery Roitblat et al. (2010). ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_2",
"text": " While Deep Learning models have achieved state-of-the-art on many NLP tasks, these models are trained from scratch, requiring large datasets, and days to converge. Research in NLP focused mostly on transductive transfer Blitzer et al. (2007). For inductive transfer, fine-tuning pretrained word embeddings Mikolov et al. (2013), a simple transfer technique that only targets a model’s first layer, has had a large impact in practice and is used in most state-of-the-art models. Recent approaches that concatenate embeddings derived from other tasks with the input at different layers Peters et al. (2017); McCann et al. (2017); Peters et al. (2018) still train the main task model from scratch and treat pretrained embeddings as fixed parameters, limiting their usefulness. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_3",
"text": " In light of the benefits of pretraining Erhan et al. (2010), we should be able to do better than randomly initializing the remaining parameters of our models. However, inductive transfer via fine-tuning has been unsuccessful for NLP Mou et al. (2016). Dai and Le (2015) first proposed fine-tuning a language model (LM) but require millions of in-domain documents to achieve good performance, which severely limits its applicability. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_4",
"text": " We show that not the idea of LM fine-tuning but our lack of knowledge of how to train them effectively has been hindering wider adoption. LMs overfit to small datasets and suffered catastrophic forgetting when fine-tuned with a classifier. Compared to CV, NLP models are typically more shallow and thus require different fine-tuning methods. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_5",
"text": " We propose a new method, Universal Language Model Fine-tuning (ULMFiT) that addresses these issues and enables robust inductive transfer learning for any NLP task, akin to fine-tuning ImageNet models: The same 3-layer LSTM architecture—with the same hyperparameters and no additions other than tuned dropout hyperparameters—outperforms highly engineered models and transfer learning approaches on six widely studied text classification tasks. On IMDb, with 100100100 labeled examples, ULMFiT matches the performance of training from scratch with 10×10\\times and—given 505050k unlabeled examples—with 100×100\\times more data. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_6",
"text": " Our contributions are the following: 1) We propose Universal Language Model Fine-tuning (ULMFiT), a method that can be used to achieve CV-like transfer learning for any task for NLP. 2) We propose discriminative fine-tuning, slanted triangular learning rates, and gradual unfreezing, novel techniques to retain previous knowledge and avoid catastrophic forgetting during fine-tuning. 3) We significantly outperform the state-of-the-art on six representative text classification datasets, with an error reduction of 18-24% on the majority of datasets. 4) We show that our method enables extremely sample-efficient transfer learning and perform an extensive ablation analysis. 5) We make the pretrained models and our code available to enable wider adoption. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_7",
"text": " Features in deep neural networks in CV have been observed to transition from general to task-specific from the first to the last layer Yosinski et al. (2014). For this reason, most work in CV focuses on transferring the first layers of the model Long et al. (2015b). Sharif Razavian et al. (2014) achieve state-of-the-art results using features of an ImageNet model as input to a simple classifier. In recent years, this approach has been superseded by fine-tuning either the last Donahue et al. (2014) or several of the last layers of a pretrained model and leaving the remaining layers frozen Long et al. (2015a). ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_8",
"text": " In NLP, only recently have methods been proposed that go beyond transferring word embeddings. The prevailing approach is to pretrain embeddings that capture additional context via other tasks. Embeddings at different levels are then used as features, concatenated either with the word embeddings or with the inputs at intermediate layers. This method is known as hypercolumns Hariharan et al. (2015) in CV333A hypercolumn at a pixel in CV is the vector of activations of all CNN units above that pixel. In analogy, a hypercolumn for a word or sentence in NLP is the concatenation of embeddings at different layers in a pretrained model. and is used by Peters et al. (2017), Peters et al. (2018), Wieting and Gimpel (2017), Conneau et al. (2017), and McCann et al. (2017) who use language modeling, paraphrasing, entailment, and Machine Translation (MT) respectively for pretraining. Specifically, Peters et al. (2018) require engineered custom architectures, while we show state-of-the-art performance with the same basic architecture across a range of tasks. In CV, hypercolumns have been nearly entirely superseded by end-to-end fine-tuning Long et al. (2015a). ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_9",
"text": " A related direction is multi-task learning (MTL) Caruana (1993). This is the approach taken by Rei (2017) and Liu et al. (2018) who add a language modeling objective to the model that is trained jointly with the main task model. MTL requires the tasks to be trained from scratch every time, which makes it inefficient and often requires careful weighting of the task-specific objective functions Chen et al. (2017). ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_10",
"text": " Fine-tuning has been used successfully to transfer between similar tasks, e.g. in QA Min et al. (2017), for distantly supervised sentiment analysis Severyn and Moschitti (2015), or MT domains Sennrich et al. (2015) but has been shown to fail between unrelated ones Mou et al. (2016). Dai and Le (2015) also fine-tune a language model, but overfit with 101010k labeled examples and require millions of in-domain documents for good performance. In contrast, ULMFiT leverages general-domain pretraining and novel fine-tuning techniques to prevent overfitting even with only 100100100 labeled examples and achieves state-of-the-art results also on small datasets. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_11",
"text": " We are interested in the most general inductive transfer learning setting for NLP Pan and Yang (2010): Given a static source task 𝒯Ssubscript𝒯𝑆\\mathcal{T}_{S} and any target task 𝒯Tsubscript𝒯𝑇\\mathcal{T}_{T} with 𝒯S≠𝒯Tsubscript𝒯𝑆subscript𝒯𝑇\\mathcal{T}_{S}\\neq\\mathcal{T}_{T}, we would like to improve performance on 𝒯Tsubscript𝒯𝑇\\mathcal{T}_{T}. Language modeling can be seen as the ideal source task and a counterpart of ImageNet for NLP: It captures many facets of language relevant for downstream tasks, such as long-term dependencies Linzen et al. (2016), hierarchical relations Gulordava et al. (2018), and sentiment Radford et al. (2017). In contrast to tasks like MT McCann et al. (2017) and entailment Conneau et al. (2017), it provides data in near-unlimited quantities for most domains and languages. Additionally, a pretrained LM can be easily adapted to the idiosyncrasies of a target task, which we show significantly improves performance (see Section 5). Moreover, language modeling already is a key component of existing tasks such as MT and dialogue modeling. Formally, language modeling induces a hypothesis space ℋℋ\\mathcal{H} that should be useful for many other NLP tasks Vapnik and Kotz (1982); Baxter (2000). ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_12",
"text": " We propose Universal Language Model Fine-tuning (ULMFiT), which pretrains a language model (LM) on a large general-domain corpus and fine-tunes it on the target task using novel techniques. The method is universal in the sense that it meets these practical criteria: 1) It works across tasks varying in document size, number, and label type; 2) it uses a single architecture and training process; 3) it requires no custom feature engineering or preprocessing; and 4) it does not require additional in-domain documents or labels. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_13",
"text": " In our experiments, we use the state-of-the-art language model AWD-LSTM Merity et al. (2017a), a regular LSTM (with no attention, short-cut connections, or other sophisticated additions) with various tuned dropout hyperparameters. Analogous to CV, we expect that downstream performance can be improved by using higher-performance language models in the future. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_14",
"text": " ULMFiT consists of the following steps, which we show in Figure 1: a) General-domain LM pretraining (§3.1); b) target task LM fine-tuning (§3.2); and c) target task classifier fine-tuning (§3.3). We discuss these in the following sections. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_15",
"text": " An ImageNet-like corpus for language should be large and capture general properties of language. We pretrain the language model on Wikitext-103 Merity et al. (2017b) consisting of 28,595 preprocessed Wikipedia articles and 103 million words. Pretraining is most beneficial for tasks with small datasets and enables generalization even with 100100100 labeled examples. We leave the exploration of more diverse pretraining corpora to future work, but expect that they would boost performance. While this stage is the most expensive, it only needs to be performed once and improves performance and convergence of downstream models. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_16",
"text": " No matter how diverse the general-domain data used for pretraining is, the data of the target task will likely come from a different distribution. We thus fine-tune the LM on data of the target task. Given a pretrained general-domain LM, this stage converges faster as it only needs to adapt to the idiosyncrasies of the target data, and it allows us to train a robust LM even for small datasets. We propose discriminative fine-tuning and slanted triangular learning rates for fine-tuning the LM, which we introduce in the following. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_17",
"text": " As different layers capture different types of information Yosinski et al. (2014), they should be fine-tuned to different extents. To this end, we propose a novel fine-tuning method, discriminative fine-tuning444 An unrelated method of the same name exists for deep Boltzmann machines Salakhutdinov and Hinton (2009).. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_18",
"text": " Instead of using the same learning rate for all layers of the model, discriminative fine-tuning allows us to tune each layer with different learning rates. For context, the regular stochastic gradient descent (SGD) update of a model’s parameters θ𝜃\\theta at time step t𝑡t looks like the following Ruder (2016): θt=θt−1−η⋅∇θJ(θ)subscript𝜃𝑡subscript𝜃𝑡1⋅𝜂subscript∇𝜃𝐽𝜃\\theta_{t}=\\theta_{t-1}-\\eta\\cdot\\nabla_{\\theta}J(\\theta) (1) where η𝜂\\eta is the learning rate and ∇θJ(θ)subscript∇𝜃𝐽𝜃\\nabla_{\\theta}J(\\theta) is the gradient with regard to the model’s objective function. For discriminative fine-tuning, we split the parameters θ𝜃\\theta into {θ1,…,θL}superscript𝜃1…superscript𝜃𝐿\\{\\theta^{1},\\ldots,\\theta^{L}\\} where θlsuperscript𝜃𝑙\\theta^{l} contains the parameters of the model at the l𝑙l-th layer and L𝐿L is the number of layers of the model. Similarly, we obtain {η1,…,ηL}superscript𝜂1…superscript𝜂𝐿\\{\\eta^{1},\\ldots,\\eta^{L}\\} where ηlsuperscript𝜂𝑙\\eta^{l} is the learning rate of the l𝑙l-th layer. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_19",
"text": " The SGD update with discriminative fine-tuning is then the following: θtl=θt−1l−ηl⋅∇θlJ(θ)superscriptsubscript𝜃𝑡𝑙superscriptsubscript𝜃𝑡1𝑙⋅superscript𝜂𝑙subscript∇superscript𝜃𝑙𝐽𝜃\\theta_{t}^{l}=\\theta_{t-1}^{l}-\\eta^{l}\\cdot\\nabla_{\\theta^{l}}J(\\theta) (2) We empirically found it to work well to first choose the learning rate ηLsuperscript𝜂𝐿\\eta^{L} of the last layer by fine-tuning only the last layer and using ηl−1=ηl/2.6superscript𝜂𝑙1superscript𝜂𝑙2.6\\eta^{l-1}=\\eta^{l}/2.6 as the learning rate for lower layers. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_20",
"text": " For adapting its parameters to task-specific features, we would like the model to quickly converge to a suitable region of the parameter space in the beginning of training and then refine its parameters. Using the same learning rate (LR) or an annealed learning rate throughout training is not the best way to achieve this behaviour. Instead, we propose slanted triangular learning rates (STLR), which first linearly increases the learning rate and then linearly decays it according to the following update schedule, which can be seen in Figure 2: cut=⌊T⋅cut_frac⌋p={t/cut,ift<cut1−t−cutcut⋅(1/cut_frac−1),otherwiseηt=ηmax⋅1+p⋅(ratio−1)ratio𝑐𝑢𝑡⋅𝑇𝑐𝑢𝑡_𝑓𝑟𝑎𝑐𝑝cases𝑡𝑐𝑢𝑡if𝑡𝑐𝑢𝑡1𝑡𝑐𝑢𝑡⋅𝑐𝑢𝑡1𝑐𝑢𝑡_𝑓𝑟𝑎𝑐1otherwisesubscript𝜂𝑡⋅subscript𝜂𝑚𝑎𝑥1⋅𝑝𝑟𝑎𝑡𝑖𝑜1𝑟𝑎𝑡𝑖𝑜\\begin{split}cut&=\\lfloor T\\cdot cut\\_frac\\rfloor\\\\ p&=\\begin{cases}t/cut,&\\text{if}\\ t<cut\\\\ 1-\\frac{t-cut}{cut\\cdot(1/cut\\_frac-1)},&\\text{otherwise}\\end{cases}\\\\ \\eta_{t}&=\\eta_{max}\\cdot\\frac{1+p\\cdot(ratio-1)}{ratio}\\end{split} (3) where T𝑇T is the number of training iterations555In other words, the number of epochs times the number of updates per epoch., cut_frac𝑐𝑢𝑡_𝑓𝑟𝑎𝑐cut\\_frac is the fraction of iterations we increase the LR, cut𝑐𝑢𝑡cut is the iteration when we switch from increasing to decreasing the LR, p𝑝p is the fraction of the number of iterations we have increased or will decrease the LR respectively, ratio𝑟𝑎𝑡𝑖𝑜ratio specifies how much smaller the lowest LR is from the maximum LR ηmaxsubscript𝜂𝑚𝑎𝑥\\eta_{max}, and ηtsubscript𝜂𝑡\\eta_{t} is the learning rate at iteration t𝑡t. We generally use cut_frac=0.1𝑐𝑢𝑡_𝑓𝑟𝑎𝑐0.1cut\\_frac=0.1, ratio=32𝑟𝑎𝑡𝑖𝑜32ratio=32 and ηmax=0.01subscript𝜂𝑚𝑎𝑥0.01\\eta_{max}=0.01. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_21",
"text": " STLR modifies triangular learning rates Smith (2017) with a short increase and a long decay period, which we found key for good performance.666We also credit personal communication with the author. In Section 5, we compare against aggressive cosine annealing, a similar schedule that has recently been used to achieve state-of-the-art performance in CV Loshchilov and Hutter (2017).777While Loshchilov and Hutter (2017) use multiple annealing cycles, we generally found one cycle to work best. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_22",
"text": " Finally, for fine-tuning the classifier, we augment the pretrained language model with two additional linear blocks. Following standard practice for CV classifiers, each block uses batch normalization Ioffe and Szegedy (2015) and dropout, with ReLU activations for the intermediate layer and a softmax activation that outputs a probability distribution over target classes at the last layer. Note that the parameters in these task-specific classifier layers are the only ones that are learned from scratch. The first linear layer takes as the input the pooled last hidden layer states. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_23",
"text": " The signal in text classification tasks is often contained in a few words, which may occur anywhere in the document. As input documents can consist of hundreds of words, information may get lost if we only consider the last hidden state of the model. For this reason, we concatenate the hidden state at the last time step 𝐡Tsubscript𝐡𝑇\\mathbf{h}_{T} of the document with both the max-pooled and the mean-pooled representation of the hidden states over as many time steps as fit in GPU memory 𝐇={𝐡1,…,𝐡T}𝐇subscript𝐡1…subscript𝐡𝑇\\mathbf{H}=\\{\\mathbf{h}_{1},\\ldots,\\mathbf{h}_{T}\\}: 𝐡c=(𝐡T,𝚖𝚊𝚡𝚙𝚘𝚘𝚕(𝐇),𝚖𝚎𝚊𝚗𝚙𝚘𝚘𝚕(𝐇))subscript𝐡𝑐subscript𝐡𝑇𝚖𝚊𝚡𝚙𝚘𝚘𝚕𝐇𝚖𝚎𝚊𝚗𝚙𝚘𝚘𝚕𝐇\\mathbf{h}_{c}=(\\mathbf{h}_{T},\\mathtt{maxpool}(\\mathbf{H}),\\mathtt{meanpool}(\\mathbf{H})) (4) where ()() is concatenation. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_24",
"text": " Fine-tuning the target classifier is the most critical part of the transfer learning method. Overly aggressive fine-tuning will cause catastrophic forgetting, eliminating the benefit of the information captured through language modeling; too cautious fine-tuning will lead to slow convergence (and resultant overfitting). Besides discriminative fine-tuning and triangular learning rates, we propose gradual unfreezing for fine-tuning the classifier. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_25",
"text": " Rather than fine-tuning all layers at once, which risks catastrophic forgetting, we propose to gradually unfreeze the model starting from the last layer as this contains the least general knowledge Yosinski et al. (2014): We first unfreeze the last layer and fine-tune all unfrozen layers for one epoch. We then unfreeze the next lower frozen layer and repeat, until we fine-tune all layers until convergence at the last iteration. This is similar to ‘chain-thaw’ Felbo et al. (2017), except that we add a layer at a time to the set of ‘thawed’ layers, rather than only training a single layer at a time. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_26",
"text": " While discriminative fine-tuning, slanted triangular learning rates, and gradual unfreezing all are beneficial on their own, we show in Section 5 that they complement each other and enable our method to perform well across diverse datasets. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_27",
"text": " Language models are trained with backpropagation through time (BPTT) to enable gradient propagation for large input sequences. In order to make fine-tuning a classifier for large documents feasible, we propose BPTT for Text Classification (BPT3C): We divide the document into fixed-length batches of size b𝑏b. At the beginning of each batch, the model is initialized with the final state of the previous batch; we keep track of the hidden states for mean and max-pooling; gradients are back-propagated to the batches whose hidden states contributed to the final prediction. In practice, we use variable length backpropagation sequences Merity et al. (2017a). ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_28",
"text": " Similar to existing work Peters et al. (2017, 2018), we are not limited to fine-tuning a unidirectional language model. For all our experiments, we pretrain both a forward and a backward LM. We fine-tune a classifier for each LM independently using BPT3C and average the classifier predictions. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_29",
"text": " While our approach is equally applicable to sequence labeling tasks, we focus on text classification tasks in this work due to their important real-world applications. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_30",
"text": " We evaluate our method on six widely-studied datasets, with varying numbers of documents and varying document length, used by state-of-the-art text classification and transfer learning approaches Johnson and Zhang (2017); McCann et al. (2017) as instances of three common text classification tasks: sentiment analysis, question classification, and topic classification. We show the statistics for each dataset and task in Table 1. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_31",
"text": " For sentiment analysis, we evaluate our approach on the binary movie review IMDb dataset Maas et al. (2011) and on the binary and five-class version of the Yelp review dataset compiled by Zhang et al. (2015). ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_32",
"text": " We use the six-class version of the small TREC dataset Voorhees and Tice (1999) dataset of open-domain, fact-based questions divided into broad semantic categories. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_33",
"text": " For topic classification, we evaluate on the large-scale AG news and DBpedia ontology datasets created by Zhang et al. (2015). ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_34",
"text": " We use the same pre-processing as in earlier work Johnson and Zhang (2017); McCann et al. (2017). In addition, to allow the language model to capture aspects that might be relevant for classification, we add special tokens for upper-case words, elongation, and repetition. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_35",
"text": " We are interested in a model that performs robustly across a diverse set of tasks. To this end, if not mentioned otherwise, we use the same set of hyperparameters across tasks, which we tune on the IMDb validation set. We use the AWD-LSTM language model Merity et al. (2017a) with an embedding size of 400400400, 333 layers, 115011501150 hidden activations per layer, and a BPTT batch size of 707070. We apply dropout of 0.40.40.4 to layers, 0.30.30.3 to RNN layers, 0.40.40.4 to input embedding layers, 0.050.050.05 to embedding layers, and weight dropout of 0.50.50.5 to the RNN hidden-to-hidden matrix. The classifier has a hidden layer of size 505050. We use Adam with β1=0.7subscript𝛽10.7\\beta_{1}=0.7 instead of the default β1=0.9subscript𝛽10.9\\beta_{1}=0.9 and β2=0.99subscript𝛽20.99\\beta_{2}=0.99, similar to Dozat and Manning (2017). We use a batch size of 646464, a base learning rate of 0.0040.0040.004 and 0.010.010.01 for fine-tuning the LM and the classifier respectively, and tune the number of epochs on the validation set of each task888On small datasets such as TREC-6, we fine-tune the LM only for 151515 epochs without overfitting, while we can fine-tune longer on larger datasets. We found 505050 epochs to be a good default for fine-tuning the classifier.. We otherwise use the same practices used in Merity et al. (2017a). ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_36",
"text": " For each task, we compare against the current state-of-the-art. For the IMDb and TREC-6 datasets, we compare against CoVe McCann et al. (2017), a state-of-the-art transfer learning method for NLP. For the AG, Yelp, and DBpedia datasets, we compare against the state-of-the-art text categorization method by Johnson and Zhang (2017). ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_37",
"text": " For consistency, we report all results as error rates (lower is better). We show the test error rates on the IMDb and TREC-6 datasets used by McCann et al. (2017) in Table 2. Our method outperforms both CoVe, a state-of-the-art transfer learning method based on hypercolumns, as well as the state-of-the-art on both datasets. On IMDb, we reduce the error dramatically by 43.9% and 22% with regard to CoVe and the state-of-the-art respectively. This is promising as the existing state-of-the-art requires complex architectures Peters et al. (2018), multiple forms of attention McCann et al. (2017) and sophisticated embedding schemes Johnson and Zhang (2016), while our method employs a regular LSTM with dropout. We note that the language model fine-tuning approach of Dai and Le (2015) only achieves an error of 7.64 vs. 4.6 for our method on IMDb, demonstrating the benefit of transferring knowledge from a large ImageNet-like corpus using our fine-tuning techniques. IMDb in particular is reflective of real-world datasets: Its documents are generally a few paragraphs long—similar to emails (e.g for legal discovery) and online comments (e.g for community management); and sentiment analysis is similar to many commercial applications, e.g. product response tracking and support email routing. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_38",
"text": " On TREC-6, our improvement—similar as the improvements of state-of-the-art approaches—is not statistically significant, due to the small size of the 500-examples test set. Nevertheless, the competitive performance on TREC-6 demonstrates that our model performs well across different dataset sizes and can deal with examples that range from single sentences—in the case of TREC-6—to several paragraphs for IMDb. Note that despite pretraining on more than two orders of magnitude less data than the 7 million sentence pairs used by McCann et al. (2017), we consistently outperform their approach on both datasets. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_39",
"text": " We show the test error rates on the larger AG, DBpedia, Yelp-bi, and Yelp-full datasets in Table 3. Our method again outperforms the state-of-the-art significantly. On AG, we observe a similarly dramatic error reduction by 23.7% compared to the state-of-the-art. On DBpedia, Yelp-bi, and Yelp-full, we reduce the error by 4.8%, 18.2%, 2.0% respectively. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_40",
"text": " In order to assess the impact of each contribution, we perform a series of analyses and ablations. We run experiments on three corpora, IMDb, TREC-6, and AG that are representative of different tasks, genres, and sizes. For all experiments, we split off 10%percent1010\\% of the training set and report error rates on this validation set with unidirectional LMs. We fine-tune the classifier for 505050 epochs and train all methods but ULMFiT with early stopping. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_41",
"text": " One of the main benefits of transfer learning is being able to train a model for a task with a small number of labels. We evaluate ULMFiT on different numbers of labeled examples in two settings: only labeled examples are used for LM fine-tuning (‘supervised’); and all task data is available and can be used to fine-tune the LM (‘semi-supervised’). We compare ULMFiT to training from scratch—which is necessary for hypercolumn-based approaches. We split off balanced fractions of the training data, keep the validation set fixed, and use the same hyperparameters as before. We show the results in Figure 3. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_42",
"text": " On IMDb and AG, supervised ULMFiT with only 100100100 labeled examples matches the performance of training from scratch with 10×10\\times and 20×20\\times more data respectively, clearly demonstrating the benefit of general-domain LM pretraining. If we allow ULMFiT to also utilize unlabeled examples (505050k for IMDb, 100100100k for AG), at 100100100 labeled examples, we match the performance of training from scratch with 50×50\\times and 100×100\\times more data on AG and IMDb respectively. On TREC-6, ULMFiT significantly improves upon training from scratch; as examples are shorter and fewer, supervised and semi-supervised ULMFiT achieve similar results. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_43",
"text": " We compare using no pretraining with pretraining on WikiText-103 Merity et al. (2017b) in Table 4. Pretraining is most useful for small and medium-sized datasets, which are most common in commercial applications. However, even for large datasets, pretraining improves performance. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_44",
"text": " In order to gauge the importance of choosing an appropriate LM, we compare a vanilla LM with the same hyperparameters without any dropout999To avoid overfitting, we only train the vanilla LM classifier for 555 epochs and keep dropout of 0.40.40.4 in the classifier. with the AWD-LSTM LM with tuned dropout parameters in Table 5. Using our fine-tuning techniques, even a regular LM reaches surprisingly good performance on the larger datasets. On the smaller TREC-6, a vanilla LM without dropout runs the risk of overfitting, which decreases performance. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_45",
"text": " We compare no fine-tuning against fine-tuning the full model Erhan et al. (2010) (‘Full’), the most commonly used fine-tuning method, with and without discriminative fine-tuning (‘Discr’) and slanted triangular learning rates (‘Stlr’) in Table 6. Fine-tuning the LM is most beneficial for larger datasets. ‘Discr’ and ‘Stlr’ improve performance across all three datasets and are necessary on the smaller TREC-6, where regular fine-tuning is not beneficial. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_46",
"text": " We compare training from scratch, fine-tuning the full model (‘Full’), only fine-tuning the last layer (‘Last’) Donahue et al. (2014), ‘Chain-thaw’ Felbo et al. (2017), and gradual unfreezing (‘Freez’). We furthermore assess the importance of discriminative fine-tuning (‘Discr’) and slanted triangular learning rates (‘Stlr’). We compare the latter to an alternative, aggressive cosine annealing schedule (‘Cos’) Loshchilov and Hutter (2017). We use a learning rate ηL=0.01superscript𝜂𝐿0.01\\eta^{L}=0.01 for ‘Discr’, learning rates of 0.0010.0010.001 and 0.00010.00010.0001 for the last and all other layers respectively for ‘Chain-thaw’ as in Felbo et al. (2017), and a learning rate of 0.0010.0010.001 otherwise. We show the results in Table 7. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_47",
"text": " Fine-tuning the classifier significantly improves over training from scratch, particularly on the small TREC-6. ‘Last’, the standard fine-tuning method in CV, severely underfits and is never able to lower the training error to 00. ‘Chain-thaw’ achieves competitive performance on the smaller datasets, but is outperformed significantly on the large AG. ‘Freez’ provides similar performance as ‘Full’. ‘Discr’ consistently boosts the performance of ‘Full’ and ‘Freez’, except for the large AG. Cosine annealing is competitive with slanted triangular learning rates on large data, but under-performs on smaller datasets. Finally, full ULMFiT classifier fine-tuning (bottom row) achieves the best performance on IMDB and TREC-6 and competitive performance on AG. Importantly, ULMFiT is the only method that shows excellent performance across the board—and is therefore the only universal method. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_48",
"text": " While our results demonstrate that how we fine-tune the classifier makes a significant difference, fine-tuning for inductive transfer is currently under-explored in NLP as it mostly has been thought to be unhelpful Mou et al. (2016). To better understand the fine-tuning behavior of our model, we compare the validation error of the classifier fine-tuned with ULMFiT and ‘Full’ during training in Figure 4. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_49",
"text": " On all datasets, fine-tuning the full model leads to the lowest error comparatively early in training, e.g. already after the first epoch on IMDb. The error then increases as the model starts to overfit and knowledge captured through pretraining is lost. In contrast, ULMFiT is more stable and suffers from no such catastrophic forgetting; performance remains similar or improves until late epochs, which shows the positive effect of the learning rate schedule. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_50",
"text": " At the cost of training a second model, ensembling the predictions of a forward and backwards LM-classifier brings a performance boost of around 0.50.50.5–0.70.70.7. On IMDb we lower the test error from 5.305.305.30 of a single model to 4.584.584.58 for the bidirectional model. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_51",
"text": " While we have shown that ULMFiT can achieve state-of-the-art performance on widely used text classification tasks, we believe that language model fine-tuning will be particularly useful in the following settings compared to existing transfer learning approaches Conneau et al. (2017); McCann et al. (2017); Peters et al. (2018): a) NLP for non-English languages, where training data for supervised pretraining tasks is scarce; b) new NLP tasks where no state-of-the-art architecture exists; and c) tasks with limited amounts of labeled data (and some amounts of unlabeled data). ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_52",
"text": " Given that transfer learning and particularly fine-tuning for NLP is under-explored, many future directions are possible. One possible direction is to improve language model pretraining and fine-tuning and make them more scalable: for ImageNet, predicting far fewer classes only incurs a small performance drop Huh et al. (2016), while recent work shows that an alignment between source and target task label sets is important Mahajan et al. (2018)—focusing on predicting a subset of words such as the most frequent ones might retain most of the performance while speeding up training. Language modeling can also be augmented with additional tasks in a multi-task learning fashion Caruana (1993) or enriched with additional supervision, e.g. syntax-sensitive dependencies Linzen et al. (2016) to create a model that is more general or better suited for certain downstream tasks, ideally in a weakly-supervised manner to retain its universal properties. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_53",
"text": " Another direction is to apply the method to novel tasks and models. While an extension to sequence labeling is straightforward, other tasks with more complex interactions such as entailment or question answering may require novel ways to pretrain and fine-tune. Finally, while we have provided a series of analyses and ablations, more studies are required to better understand what knowledge a pretrained language model captures, how this changes during fine-tuning, and what information different tasks require. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
},
{
"id": "1801.06146_all_54",
"text": " We have proposed ULMFiT, an effective and extremely sample-efficient transfer learning method that can be applied to any NLP task. We have also proposed several novel fine-tuning techniques that in conjunction prevent catastrophic forgetting and enable robust learning across a diverse range of tasks. Our method significantly outperformed existing transfer learning techniques and the state-of-the-art on six representative text classification tasks. We hope that our results will catalyze new developments in transfer learning for NLP. ",
"title": "Universal Language Model Fine-tuning for Text Classification"
}
] |
Why should LSTM based auto-encoder models learn good features?
| "Since LSTM based auto-encoder models control the learning an identity mapping, it forced learn good(...TRUNCATED)
|
[
14
] | [{"id":"1502.04681_all_0","text":" Understanding temporal sequences is important for solving many pr(...TRUNCATED)
|
"Can the the NetVLAD pooling layer be inserted into any other CNN or does it support certain archite(...TRUNCATED)
|
Yes, it is a generic building block and can be inserted into any other CNN architectures [48].
|
[
48
] | [{"id":"1511.07247_all_0","text":" Visual place recognition has received a significant amount of att(...TRUNCATED)
|
"Why does author experiment the quantized linear supernet design even though Radosavovic et al. alre(...TRUNCATED)
| "The previous study shows the linear design is beneficial in terms of computational complexity, whil(...TRUNCATED)
|
[
34
] | [{"id":"2009.02009_all_0","text":" As there are growing needs of deep learning applications based on(...TRUNCATED)
|
"Why did the dot scoring function perform well for global attention while the general scoring functi(...TRUNCATED)
| "Authors thinks it's interesting to observe that \"dot\" works well for the global attention and \"g(...TRUNCATED)
|
[
38
] | [{"id":"1508.04025_all_0","text":" Neural Machine Translation (NMT) achieved state-of-the-art perfor(...TRUNCATED)
|
Is it true that the first-order approximation led to roughly 33% speed-up in network computation?
| "According to the paper, eliminating the Hessian calculation increases the overall calculation speed(...TRUNCATED)
|
[
31
] | [{"id":"1703.03400_all_0","text":" Learning quickly is a hallmark of human intelligence, whether it (...TRUNCATED)
|
"Why didn't the authors intend a \"chord\" to represent a more meaningful unit in music, such as a b(...TRUNCATED)
| "The authors intend a \"chord\" to represent simultaneous notes to intuitively models a polyphonic s(...TRUNCATED)
|
[
5,
6
] | [{"id":"2208.14867_all_0","text":" Computational modeling of expressive music performance focuses on(...TRUNCATED)
|
Who collected the queries from MSMARCO-Passage dataset to make MSMARCO-TRAIN query set?
| "MARCO-Passage collection is a large-scale publicly available corpus and two query sets derived from(...TRUNCATED)
|
[
40
] | [{"id":"2204.11673_all_0","text":" Passage Re-ranking is a crucial stage in modern information retri(...TRUNCATED)
|
End of preview. Expand
in Data Studio
README.md exists but content is empty.
- Downloads last month
- 12