Fish RNNs learn compact codes that better decode homing angles and social distance
tweetsatpreet · x · 2026-07-29
- The RNN activity has a much lower effective rank than network width, suggesting the agents use a compact internal representation.
- It robustly encodes task and social variables: for example, the field-line homing angle decodes better than the naive straight-line angle.
- Two fish align their neural activity more at short range than at long range.
More from Research
- DeepMind’s AI co-scientist solved a 10-year bacterial gene-transfer problem in two days — nathanbenaich · 2026-07-30
- AI Moves Too Fast: NAACL 2027 Workshop Proposals Slammed as Outdated — yanaiela · 2026-07-30
- Deep Dive: Will AI Make Formal Verification Mainstream? — The Pragmatic Engineer · 2026-07-30
- Four months of Gabor-wavelet image generation led to a broader Claude-assisted research project — pixlpa · 2026-07-30
- Agent Retrieval Bench shows coding agents often fail before patch generation even starts — Bowen Qin · 2026-07-30
- When MCP servers drift, user settings may stop applying to the capability you meant — Loocor · 2026-07-30