3. Predictive processing / Free Energy Principle (FEP) / active inference

#3.1 The canonical claim and its critics

Colombo & Wright, *First principles in the life sciences: the free-energy principle, organicism, and mechanism, Synthese* (2018/2021), doi 10.1007/s11229-018-01932-w. Abstract verbatim (this is the central falsifiability indictment in the literature):

"The free-energy principle states that all systems that minimize their free energy resist a tendency to physical disintegration. Originally proposed to account for perception, learning, and action, the free-energy principle has been applied to the evolution, development, morphology, anatomy and function of the brain, and has been called a postulate, an unfalsifiable natural law or an imperative. While it might afford a theoretical foundation for understanding the relationship between environment, life, and mind, its epistemic status is unclear. Also unclear is how the free-energy principle relates to prominent theoretical approaches to life science phenomena, such as organicism and mechanism. This paper clarifies both issues, and identifies limits and prospects for the free-energy principle as a first principle in the life sciences."

Bruineberg, Dołęga, Dewhurst, Baltieri, The Emperor's New Markov Blankets, *Behavioral and Brain Sciences* (2021/2022), doi 10.1017/s0140525x21002351. Abstract verbatim:

"…we identify a persistent confusion in the literature between the formal use of Markov blankets as an epistemic tool for Bayesian inference, and their novel metaphysical use in the free energy framework to demarcate the physical boundary between an agent and its environment. Consequently, we propose to distinguish 'Pearl blankets' to refer to the original epistemic use of Markov blankets and 'Friston blankets' to refer to this new metaphysical construct. Second, we use this distinction to critically assess claims resting on the philosophical application of Markov blankets… We suggest that the literature would do well in differentiating two different research programmes: 'inference as model' and 'inference as world model.' Only the latter is capable of doing philosophical work, but it requires additional premises that cannot be justified by appeal to the success of the mathematical framework alone."

Aguilera, Millidge, Tschantz, Buckley, How particular is the physics of the free energy principle? — the technical critique that elicited four published replies. I verified it only indirectly, via the titles of the replies: arXiv 2205.07793 (*Regarding Flows Under the Free Energy Principle: A Comment on "How Particular is the Physics of the Free Energy Principle?"*), arXiv 2204.13576, and arXiv 2205.10190. I did not verify its own arXiv ID, so I do not state one.

Biehl, Pollock, Kanai, A Technical Critique of Some Parts of the Free Energy Principle, arXiv 2001.06408 (Entropy). Verbatim, and note this is a technical, not philosophical, attack:

"We prove by counterexample that the original free energy lemma, when taken at face value, is wrong… We show that crucial steps in the free energy argument which involve rewriting the equations of motion of systems with Markov blankets, are not generally correct without additional (previously unstated) assumptions… We show further that this free energy lemma, when it does hold, implies equality of variational density and ergodic conditional density. The interpretation in terms of Bayesian inference hinges on this point, and we hence conclude that it is not sufficiently justified."

Friston, Da Costa, Parr, Some interesting observations on the free energy principle, arXiv 2002.04501 (Entropy 23:1076, 2021) — the reply, focused on *"solenoidal coupling between various (subsets of) states in sparsely coupled systems that possess a Markov blanket — and the distinction between exact and approximate Bayesian inference."*

Defenders' partial concessions are worth noting. Heins & Da Costa, arXiv 2205.10190: *"The authors demonstrate that in general, Markov blankets are not guaranteed to follow from sparse coupling. The current commentary explains the relationship between sparse coupling and Markov blankets in the case of Gaussian steady-state densities… Future work should focus on verifying whether these sorts of constraints are satisfied in realistic models of sparsely coupled systems." And Sakthivadivel, arXiv 2205.07793: "this piece takes the position that the application of a state-based formulation of the FEP is inappropriate for certain simple systems, but, that the FEP can be expected to hold regardless." — i.e. the FEP is defended by narrowing its testable content*.

Against Bruineberg's sharp instrumental/ontological dichotomy: Seth, Korbak, Tschantz, arXiv 2201.06900: *"proposing a sharp distinction neglects the value of recognising a continuum spanning from instrumental to ontological."*

⚠️ I could not verify a paper titled "The Free Energy Principle: A Critique" by Colombo & Wright. What exists (verified) is the Synthese paper above. I am flagging this because your brief named a title that I could not confirm.

#3.2 What active inference actually gives you that is testable

This is the important part for Darlin. Active inference, unlike the FEP-as-metaphysics, has hard, falsifiable content when instantiated:

• Da Costa, Sajid, Parr, Friston, Smith, Reward Maximisation through Discrete Active Inference, arXiv 2009.08111. This is the crispest testable claim in the whole FEP literature, verbatim: *"we show the conditions under which active inference produces the optimal solution to the Bellman equation… On partially observed Markov decision processes, the standard active inference scheme can produce Bellman optimal actions for planning horizons of 1, but not beyond. In contrast, a recently developed recursive active inference scheme (sophisticated inference) can produce Bellman optimal actions on any finite temporal horizon."* → This is a sharp, falsifiable prediction: standard active inference should match optimal control at horizon 1 and systematically fail beyond it; sophisticated inference should not. A small RL agent can test this directly against a known-optimal POMDP.

• Tschantz, Millidge, Seth, Buckley, Reinforcement Learning through Active Inference, arXiv 2002.12636. Claim: *"we develop and implement a novel objective for decision making, which we term the free energy of the expected future. We demonstrate that the resulting algorithm successfully balances exploration and exploitation, simultaneously achieving robust performance on several challenging RL benchmarks with sparse, well-shaped, and no rewards."* → testable: performance with the reward channel removed entirely. This is a strong, unusual prediction.

• Tschantz, Baltieri, Seth, Buckley, Scaling active inference, arXiv 1911.10601: *"proof-of-principle results demonstrating efficient exploration and an order of magnitude increase in sample efficiency over strong model-free baselines."*

• van der Himst & Lanillos, Deep Active Inference for Partially Observable MDPs, arXiv 2009.03622: *"our approach has comparable or better performance than deep Q-learning"* on OpenAI benchmarks, learning from high-dimensional pixels by optimising a variant of expected free energy with a VAE state representation.

• Noel, van Hoof, Millidge, *Online reinforcement learning with sparse rewards through an active inference capsule, arXiv 2106.02390: reports "robustness to observation noise, which in fact improves performance"* — a counter-intuitive, testable prediction.

• Millidge, Deep Active Inference as Variational Policy Gradients, arXiv 1907.03876: *"our algorithm shows similarities with maximum entropy reinforcement learning and the policy gradients algorithm."* → important negative result for anyone claiming active inference is a different kind of thing: at scale it is close to MaxEnt RL.

• Schneider, Belousov, Abdulsamad, Peters, Active Inference for Robotic Manipulation, arXiv 2206.10313: *"we conclude that using an information-seeking objective is beneficial in sparse environments and allows the agent to solve tasks in which methods that do not exhibit directed exploration fail."*

• Mazzaglia, Verbelen, Dhoedt, Contrastive Active Inference, arXiv 2110.10083 (NeurIPS 2021): *"our approach closely matches their performance" against RL agents with* hand-designed rewards. → i.e. active inference does not beat reward-engineered RL here.

★ Assessment of FEP/active inference. Split the claim in two.

*The FEP as a universal principle about all self-organising systems is unfalsifiable and should not be built on.* Colombo & Wright state this directly ("unfalsifiable natural law or imperative"); Biehl, Pollock & Kanai prove the central lemma "is wrong" as stated; Bruineberg et al. show the Markov-blanket move requires premises "that cannot be justified by appeal to the success of the mathematical framework alone"; and the defenders' own replies concede the formulation must be narrowed for "certain simple systems" while asserting it "can be expected to hold regardless" — which is the signature of an unfalsifiable framework.

*Active inference as a specific algorithm with expected-free-energy objectives is falsifiable and already tested.* The concrete, citable testable claims are: horizon-1-only Bellman optimality for standard schemes; reward-free operation; noise robustness; sparse-reward sample efficiency. Note two cautions: Millidge's result that deep active inference approximates MaxEnt policy gradients (so "active inference" may be relabelled RL), and Mazzaglia et al.'s result that it only matches well-tuned RL. For Darlin: use active-inference-style epistemic objectives as an engineering choice with measurable benchmarks — never as evidence of consciousness.