New paper: population codes reveal more interpretable visual features than single neurons
Hidenori8Tanaka · x · 2026-09-12
A new bioRxiv paper from Cold Spring Harbor Laboratory (Habon Issa, Sunny Liu, Jona Ballé, David Klindt) challenges Barlow's 1972 neuron doctrine framework of finding per-neuron 'trigger features':
- Method: extract population codes from macaque V4/IT electrophysiology datasets and vision models, plus an automated method to measure representation interpretability and diversity.
- Findings: populations represent more interpretable and diverse features than single neurons, with the advantage growing as sampled neuron count and image diversity increase.
- Targeted ablations demonstrate the functional role of interpretable representations.
- The work resonates with the superposition concept from AI interpretability, informing both neuroscience and machine learning.
More from Models
- GPT-Live τ-Voice scores span 20 points across sources; backend model choice matters — rdesh26 · 2026-09-12
- GPT-Live's $0.05/min pricing works out to ~$28/M tokens, hints at mid-size frontend model — rdesh26 · 2026-09-12
- Devs reverse engineer Astra's computer use and make it 2x faster — iamrobotbear · 2026-09-12
- DeepSeek ships V4.1-Flash with native vision; builds landing pages for 1/40th of Claude's cost — togethercompute · 2026-09-12
- Abliterated GLM 5.3 goes viral, but the technique is 3 years old — thursdai_pod · 2026-09-12
- A 671B model only activates ~37B per token: how MoE works and what it costs — techNmak · 2026-09-12