GitHub

Active inference is the claim that perception, action, and learning are all one process. An agent perceives and acts so as to minimize surprise, the gap between what it senses and what its model of the world predicted. Over about two decades Karl Friston and others have stretched this one idea to cover the cortex, psychiatric symptoms, robot control, and reinforcement learning.

Diagram showing the generative process and generative model coupled through observations and actions in a perception-action cycle
The whole setup in one picture. The world and the agent's model of it touch at only two points, what the agent observes and what it does. Everything in between is inference. From Smith, Friston & Whyte, J. Math. Psychol. 2022 (CC BY-NC-ND 4.0).

I think it is one of the most interesting ideas in neuroscience, and one of the hardest to actually learn. The papers run across twenty years and several fields that barely cite each other, the math is heavy, and most introductions assume you have already read the other introductions. I bounced off it a few times before it stuck. Awesome Active Inference is the reading list I wish someone had handed me at the start. It runs from textbooks and tutorials through the original free-energy papers, predictive coding, the discrete and continuous-time formulations, deep active inference, computational psychiatry, and the software libraries, in roughly the order I would read them now. There are two starting points at the top. Neuroscientists and machine-learning people need almost entirely different first papers, and the list says which.

The fastest way in for me was predictive coding, the idea that the cortex spends most of its effort predicting its own inputs and passing the errors upward. Rafal Bogacz’s 2017 tutorial writes this out as a small circuit you can actually simulate.

Schematic of a predictive coding neural network with prediction-error nodes and inhibitory and excitatory connections
Predictive coding as an actual circuit. Nodes send predictions down and prediction errors (ε) back up, using only local connections. From Bogacz, J. Math. Psychol. 2017 (CC BY 4.0).

The other half is decision-making. In the discrete-state version an agent plans over a POMDP and trades off getting reward against getting information. Exploration drops out of the same quantity the agent is already minimizing. The continuous-time version connects to control theory and robotics. The same machinery serves as both a theory of what brains do and a recipe for building agents.

Explore-exploit slot machine task framed as a POMDP with hidden state factors and outcome modalities
A toy task from Smith et al.'s tutorial. Two slot machines, a hint you can pay for, and a choice about whether the information is worth the cost. This is what a POMDP looks like once it is concrete. From Smith, Friston & Whyte, J. Math. Psychol. 2022 (CC BY-NC-ND 4.0).

Like the Awesome Computational Primatology list, this is a public GitHub repo. If some paper, tutorial, lecture, or library is the thing that made active inference click for you, please add it.