Ezhuthu-JEPA
Seam-Masked Latent JEPA for Tamil Akshara Recognition — A Cheap-Baseline Falsification
What It Is
A concluded negative-results study. It tested whether masking a Tamil akshara at its grapheme seam (the base-consonant / vowel-sign boundary) and predicting the masked region's latent (I-JEPA style) teaches a model the abugida's base × sign composition better than generic masking or pixel reconstruction — and thereby recognises rare, low-frequency compounds more accurately. On rendered multi-font Tamil at evenings-scale (ViT-Tiny/8, ~20k augmented instances, one RTX 5090), the hypothesis was falsified on both of its pre-registered kill criteria. The honest negative is the contribution.
This is the cheap-baseline-falsification rule doing exactly its job. The mandated baselines (block-masking JEPA and MAE pixel-target) were run on the paper's own metric and beat the mechanism before any expensive scaling — so ~2 GPU-hours and one evening replaced a full multi-seed sweep, a degradation suite, and a paper built on an effect that isn't there. Publishing the negative, with the benchmark and harness that produced it, ends the project at the right time.
What I Built
- Pre-registered the null hypothesis and kill criteria — ε = 2.0 pp on metric M (bottom-quartile-frequency akshara top-1 accuracy), adjudicated by non-overlapping bootstrap CIs and a paired McNemar test — before any arm ran
- Frequency-stratified Tamil akshara benchmark from a 172-work Project Madurai corpus (4.85M uyirmei counted) with a font-holdout augmented split, tightening the metric-M CI to ~1 pp
- One switchable ViT-Tiny/8 pretraining loop with a single changed variable per arm — seam-JEPA (latent), block-JEPA (random-location control), and MAE-at-seam (pixel-target control), compute and mask-ratio held exactly fixed
- Encoder-agnostic evaluation harness (metric M, bootstrap CIs, seam-source / font stratification, paired McNemar comparator) and a multi-font HarfBuzz/FreeType rendering + grapheme-seam-labelling pipeline
- Five-identifier provenance/config contract (config hash, code SHA, data hash, seed, environment) recorded before every run
Tech Stack
Concluded — hypothesis falsified on both kill criteria (DEC-0019, G1 = BLOCK). Block-masking JEPA (M = 0.335) beat seam-masking JEPA (0.290) past the pre-registered ε with non-overlapping CIs; MAE-at-seam / pixel target (0.532) beat the latent target by ~24 pp; both latent arms fell below a raw-pixel baseline (0.359) and degraded with more training. The base→sign premise probe cleared chance but was dominated by sign location, not base-ink composition. Cost to falsify: ~2 GPU-hours. Benchmark, evaluation harness, rendering pipeline, and pretraining loop released. github.com/Murai-Labs/Ezhuthu-Jepa.