NeuroAI researcher: backprop conflates evolution, development and lifelong learning
ShahabBakht · x · 2026-09-15
In a neuroAI thread, the author makes two critiques:
- Three regimes, one algorithm: evolution, development and lifelong learning operate on completely different scales and constraints, but neuroAI often blurs them — and backprop lumps all three stages together, which is both the strength and weakness of current learning regimes.
- ImageNet isn't the brain's target: large-scale benchmarks aren't realistic targets for biological learning, even in primates. ImageNet-trained models share representational similarities with parts of the visual hierarchy, but the brain almost certainly didn't acquire them via supervised (or self-supervised) object categorization — the similarity is an algorithmic/computational finding, not a mechanistic one.
More from Research
- Dev's CPU-native LLM Architecture Hits 113-130 tok/s on a 10B Model, Quality Lags — WildPino25 · 2026-09-15
- Anthropic paper reframes double descent as a phase transition from memorization to structure — gordic_aleksa · 2026-09-15
- Mech Interp Finding: Double Descent Is a Phase Transition Between Memorization and Generalization — gordic_aleksa · 2026-09-15
- New paper benchmarks LLM professional knowledge by turning authoritative sources into test questions — RishiBommasani · 2026-09-15
- TabPFN-3.5 launches with SOTA on complex tabular data, up to 6x faster inference — FrankRHutter · 2026-09-15
- Google's AI-in-science study mines 15M Gemini interactions, 2,600 models and 600-scientist survey — danielrock · 2026-09-15