Study: Evaluation Resolution Creates Artifacts in Brain-Model Comparisons
ConfusionSpiritual19 · reddit · 2026-08-22
A preprint study suggests that the widespread claim that untrained CNNs match or surpass backprop-trained CNNs in the early visual cortex (V1) is largely an artifact of evaluation resolution. The study shows that the gap between trained and untrained models exhibits a non-monotonic trend across resolutions from 32px to 224px. Only at higher resolutions do trained models show a clear advantage. The study also fixes a batch-norm evaluation mode bug found in earlier preprints.
More from Research
- New Architecture RHEA: Train 1B Model on 8GB VRAM — zemondza · 2026-08-24
- Trained two 16M-param models to do generative CAD with real physics — debreuil · 2026-08-24
- Claude model helps discover complex structure on S^6, solving 60-year-old math problem — Singularitarian · 2026-08-24
- Study: Agents read instructions/notes 60.5% of the time, rarely touch API docs — dair_ai · 2026-08-24
- Claude Verifies 43 Lean Modules autonomously, Tackling Theoretical Physics — Tkaraletsos · 2026-08-24
- AI fakes memory: why it gets confidently wrong without forgetting — PrajwalTomar_ · 2026-08-24