Cosyne 2026

Single consolidated note for the only Cosyne attended so far.

Conference leads and follow-ups

1. Joshua Dudman — dopamine learning signal

   • Nature paper: https://www.nature.com/articles/s41586-022-05614-z

2. Francois Rivest (Canada) — timing using DDM

   • Google Scholar: https://scholar.google.com/citations?hl=en&user=lb1fNb0AAAAJ&view_op=list_works&sortby=pubdate

3. Exponential and accumulation are same?

   • Look for relevant paper(s) by this author: https://scholar.google.com/citations?user=lMS8agIAAAAJ&hl=en

4. Uchida — TD value calculation circuit

   • bioRxiv preprint: https://www.biorxiv.org/content/10.1101/2025.09.18.677203v1

5. Michael Leprori — Brown University

   • Contravariance principle
   • Harder tasks → more unique solution

6. Cell learning / single-cell “working memory”

   • Aneta Koseska
   • eLife paper: https://elifesciences.org/articles/76825
   • Note: working memory in cells
   • ChatGPT explanation / summary link: https://chatgpt.com/s/t_69bad2ddfad081919e040b58a40c292e
   • Skeptical note: another example of overclaiming single-cell behaviour. They frame it as working memory, but it may just be slow decay of the signal caused by underlying biochemistry — basically the same flavor as habituation, i.e. decay of a remanent signal.

7. Iain M. Banks

   • Culture series
   • Note: often referred to informally here as “Ian Banks”

8. Halstead complexity

   • Constraint on AlphaEvolve’s solutions
   • https://en.wikipedia.org/wiki/Halstead_complexity_measures

9. Barlow vs Hebb

   • Find/reference the relevant paper

Related talks and references

COSYNE 2024 talks

• Playlist: https://www.youtube.com/playlist?list=PL9YzmV9joj3EjkmmUEodJNDq9ekI7iFjq
• Session 3
  • What is intelligence — life and prediction are safe. How order emerges out of chaos.
  • Estrogen regulates dopamine and enhances learning by suppressing re-uptake of dopamine.
  • Ching Fang, Abbott’s lab — adding auxiliary loss helps better learning; multi-region modelling using Deep RL.
  • Christopher Zimmerman — learning from events that happened well in the past (hours before); shows how a brain region is identified.
  • Neural coding.
  • Srjan Osdac — geometry of responses in IC, A1; how manifolds (PC-1,2,3) change with time (100 ms blocks).

Interpretability

• Belief dynamics
  • Paper: https://arxiv.org/abs/2511.00617
• Rabbit hull paper
  • Task geometry paper
• Marginal value theorem
  • For foraging; visiting time is proportional to reward rate
• Mechanical problem solving in mice
  • Task where mice have to do lever presses / slide things to get reward
  • Potentially a good paper for compositionality
• Dragon king theory
• BBP phase transition
  • The BBP phase transition (named after Jinho Baik, Gérard Ben Arous, and Sandrine Péché) describes a phenomenon in Random Matrix Theory where the largest eigenvalue of a “spiked” random matrix suddenly detaches from the main bulk of eigenvalues once the strength of a signal exceeds a critical threshold.
  • Usage note: neural network training — analyze the Hessian (curvature) of the loss landscape at initialization to see whether a gradient-based method can “find” the signal.