The Free Energy Principle (FEP) is a principle that suggests that dynamic systems, including the brain and other physical systems, are organized to minimize prediction errors, or the difference between the predictions made about the environment and the actual outcomes experienced. According to the FEP, dynamic systems encode information about their environment in a way to reduce surprisal from its input. The FEP proposes that dynamic systems are motivated to minimize prediction errors in order to maintain stability within the environment. FEP has been influential in neuroscience and neuropsychology and more recently has been used to describe systems on all spatiotemporal scales, from cells and biological species to AIs and societies.
FEP gives rise to Active Inference: a process theory of agency, that can be seen both as an explanatory theory and as an agent architecture. In the latter sense, Active Inference rivals Reinforcement Learning. It has been argued that Active Inference as an agent architecture manages the model complexity (i. e., the bias-variance tradeoff) and the exploration-exploitation tradeoff in a principled way, favours explicit, disentangled, and hence more interpretable belief representations, and is amenable for working within hierarchical systems of collective intelligence (which are seen as Active Inference agents themselves). Building ecosystems of hierarchical collective intelligence can be seen as a proposed solution for and an alternative conceptualisation of the general problem of alignment.
FEP/Active Inference is an energy-based model of intelligence: a FEP agent minimises an informational quantity called variational free energy (VFE), and Active Inference nuances this picture further, modelling agents as minimising an informational quantity called expected free energy (EFE), which is derived from VFE. This likens FEP/Active Inference to Bengio’s GFlowNets and LeCun’s Joint Embedding Predictive Architecture (JEPA), which are also energy-based. On the other hand, this distinguishes FEP/Active Inference from Reinforcement Learning, which is a reward-based model of agency, and, more generally, utility-maximising decision theories.
Active Inference is one of the most general theories of agency. It can be seen as a generalisation of the predictive coding theory of brain function (or, the Bayesian Brain hypothesis). Specifically, while predictive coding explains the agent’s perception as Bayesian inference, Active Inference models both prediction and action as inference under the single unifying objective: minimisation of the agent’s VFE or EFE. Active Inference also recovers Bayes-optimal reinforcement learning, optimal control theory, and Bayesian Decision Theory (aka EDT) under different simplifying assumptions.
The mathematical content of Active Inference is based on Variational Bayesian methods.
Parr, Thomas, Giovanni Pezzulo, and Karl J. Friston. Active inference: the free energy principle in mind, brain, and behavior. MIT Press, 2022.
Friston, Karl J., Maxwell JD Ramstead, Alex B. Kiefer, Alexander Tschantz, Christopher L. Buckley, Mahault Albarracin, Riddhi J. Pitliya et al. “Designing Ecosystems of Intelligence from First Principles.” arXiv preprint arXiv:2212.01354 (2022).
Kaufmann, Rafael, Pranav Gupta, and Jacob Taylor. “An active inference model of collective intelligence.” Entropy 23, no. 7 (2021): 830.
Bengio, Yoshua. “GFlowNet Tutorial.” (2022).
LeCun, Yann. “A path towards autonomous machine intelligence.” preprint posted on openreview (2022).
Friston, Karl, Lancelot Da Costa, Danijar Hafner, Casper Hesp, and Thomas Parr. “Sophisticated inference.” Neural Computation 33, no. 3 (2021): 713-763.