Overview
Transcript of a three-session series of discussions of the paper "Canonical neural networks perform active inference" by Takuya Isomura, Hideaki Shimazaki & Karl J. Friston. https://www.nature.com/articles/s42003-021-02994-2 LS #051.0: Background and context. https://www.youtube.com/watch?v=ZASG-rtkXDk LS #051.1: First participatory group discussion. https://www.youtube.com/watch?v=IM_NlUzyq8M LS #051.2: Second participatory group discussion. https://www.youtube.com/watch?v=hY_CajLpt9Q
Overview source: Curated paper metadata.
Methods and contributions
Read the source for the full argument, qualifications, and evidence.
Findings and contributions
- As read in the stream, the paper claims a class of canonical rate-coding neural networks is universally characterised as variational Bayesian inference under a particular but generic generative model.
- Friedman relays the paper's corollary of the complete class theorem: any neural network minimizing a cost function can be viewed as variational Bayesian inference under some priors.
- Quoted from the paper: a network of rate coding models with sigmoid activation formally corresponds to a class of POMDP models.
- The paper's maze-task simulations are described as supporting the proposition, and the implicit prior was estimated from simulated neural activity as a good approximation to the true value.
Methods
- Three-session Active Inference livestream (#051) on Isomura et al. 2022 — The work is a transcript of livestream #051: a solo background session (#051.0) plus two participatory discussions with Takuya Isomura.
- Solo slide walkthrough by Daniel Friedman with colour-coded paper quotes — In #051.0 Friedman walks through the paper's structure on slides, colour-coding direct quotes by perception, action, and complete class theorem themes.
- Discussion of the complete class theorem via external online resources — Friedman introduces the complete class theorem using external blog/LessWrong posts, inviting others with more expertise to contribute.
Summary sources: Paper metadata and evidence · Extracted source text.
Citation
Citation metadata follows the unified bibliography.
Catalog details and resources
Related in Active Inference
Other catalogued works in the same domain.