Pras13/latent2rgb-ebid-results

latent2rgb EBID experiment results

Derived results (metrics/statistics, no raw video) from EBID (Entropy-Based Instability Detection) experiments on V-JEPA2 rollouts, part of latent2rgb, issue #1 (experiments specified by Hussain Ather, pcc collaboration).

Model: V-JEPA2 ViT-L (frozen, no fine-tuning). Source clips: a subset of Something-Something V2 (ssv2) and Kinetics (kinetics_mini) — clip identifiers are included in the CSVs for traceability, but the video content itself is not redistributed here; fetch it from the original datasets under their own licenses.

Caveat carried over from the source repo: V-JEPA2's predictor was pretrained only for masked spatiotemporal completion within a fully-observed clip, never on a strictly-causal past-only-context / future-only-target split. Every result here queries it in that causal (out-of-distribution) regime.

Files

experiment1_results.csv

Per-(clip, horizon k) row. One "ensemble" = 6 Monte Carlo members (context tokens perturbed with ε=0.05 Gaussian noise); member 0 is the unperturbed rollout.

column	meaning
dataset, clip_id, t, k	sample identity; k = rollout horizon in frames
latent_l2_drift	raw L2 distance, predicted vs. true latent
cosine_distance	1 - cosine similarity, predicted vs. true latent
normalized_l2	relative L2 error (stands in for the spec's ambiguous "normalized L3")
ensemble_variance	mean per-dimension variance across the 6 ensemble members
effective_rank	exp(spectral entropy) of the ensemble spread
mahalanobis	distance of the true latent from the ensemble mean, in the ensemble's own covariance metric
ebid_entropy	spectral entropy of the ensemble spread — the EBID candidate signal
pixel_error	decoded-frame pixel error vs. ground truth — the failure signal everything else is scored against

experiment1_summary.json

Per-horizon means of the four headline signals, and each candidate's Pearson r against pixel_error, pooled across all (clip, k) samples and ranked by |r|.

experiment3_results.csv / experiment3_summary.json

Per-clip lead-time analysis: does the EBID entropy-rate drop before pixel_error crosses a crash threshold (75th percentile, pooled)? Uses the default PCC-ported threshold (2 consecutive points below the trace's own 10th-percentile entropy-rate). Result: 0/24 clips produced a warning under this threshold — attributed to the 5-point horizon grid (k=2,4,8,16,32) being too sparse for the 2-consecutive-point criterion, not a negative result on EBID itself. See entropy_threshold_used per clip for the computed threshold and crash_k for where (if at all) each clip crossed the pixel-error threshold.

experiment4_results.csv / experiment4_summary.json

Full factorial sweep: context-correction mode (pure vs corrected — decode/re-encode round trip at a fixed reinjection point) × perturbation epsilon (0.0–0.8, Gaussian noise on the reinjected tokens) × downstream horizon j (2–32 frames after reinjection). 26 clips × 96 conditions = 2496 rows. Run on real V-JEPA2 ViT-L on GPU (AWS g6.xlarge, L4).

experiment5_summary.json

Interaction/regime-boundary analysis of Experiment 4: the context-correction mode's pixel_error advantage (mode_gap) by (epsilon, j), and the entropy-vs-epsilon linear fit (slope, R²) by (mode, j).

experiment6_summary.json

Non-transitivity test: per-clip majority-vote pairwise comparison (pixel_error) across all 96 Experiment 4 conditions, searched for 3-cycles (A beats B beats C beats A), plus a Copeland-ranking consistency score.

Headline findings so far

• Ensemble spread (ebid_entropy, effective_rank) does not grow monotonically with horizon in this run.
• ebid_entropy (r=0.106) is competitive with but not clearly ahead of the other baselines at predicting pixel_error; normalized_l2 (r=0.121) is marginally stronger. latent_l2_drift is the weakest (r=0.013).
• No early-warning signal detected at the current horizon resolution (Experiment 3) — motivated the denser horizon grid used in Experiment 4.
• Experiment 4: epsilon barely moves entropy or pixel_error at all (means flat to 3 decimals across 0.0–0.8). Context-correction has a small, consistent effect: lower pixel_error (0.279 vs ~0.282) but higher entropy (2.17 vs ~2.13) than the uncorrected rollout.
• Experiment 5: the correction mode's pixel_error advantage never flips sign across (epsilon, j) — no regime boundary there. But the entropy-vs-epsilon slope flips sign between modes (negative for pure, positive and much noisier for corrected) — an interaction, not a boundary.
• Experiment 6: real non-transitive cycles exist in the per-clip pairwise comparison across sweep conditions (found, not just possible), and Copeland-ranking consistency is only ~17% of decided pairs — far from a total order. Read alongside Experiment 4's near-flat means, this looks like the conditions are largely indistinguishable in aggregate, so per-clip comparisons are noise-dominated — which is exactly what produces abundant cycles, not necessarily evidence of rich non-transitive dynamics on its own.

Raw numbers only; no causal claims beyond what's stated above.