Do Markov Blankets Give Us Real Neural Parts?

In my last post, I laid out what I think the rules are for developing a mechanistic model of the neural scale contribution to behaviour. I ended there with a question: what counts as a real neural part? How can we successfully decompose neural activity supporting a given perception-action loop into parts? 

In this post, I want to discuss one potential option: the hypothesis that Markov blankets, a key feature of the free-energy principle approach to neuroscience, can identify and pick out real neural parts. I'll discuss some recent ecological critiques of Markov blankets and some potential answers to the challenges.

Markov Blankets

Markov blankets are a formalism of the boundary of a state of a system. The boundary for a given state S comprises states that directly affect S, states directly affected by S, and states directly affecting states directly affected by S. This boundary effectively picks out what in a system S is part of, and what it is not part of. 

In practice, this blanket is identified from an adjacency matrix, in which all the parts of the system are listed down the columns and across the rows. Each cell represents the degree with which these parts covary in their activity; a 0 in the cell means those parts are statistically independent of each other. 

So Markov blankets seem to be about parts of systems; about finding which pieces work as pieces and which work separately. 

The Emperor's New Markov Blanket Trick

Two recent papers have raised concerns about the relationship between Markov blankets and real parts. Both hinge on an issue raised by Mel Andrews in their paper 'The Math is Not the Territory'. Andrews sensibly notes that there is the FEP model, and there are the system being modelled by the FEP. It's important to identify whether a feature of an FEP application is a feature of the model, or a feature of the system being modelled, and conflating these is an error. 

The first critique comes from Bruineberg et al (2021). Like Andrews, they describe two different uses of the Markov blanket concept. The first (which they call Pearl blankets, after the person who developed them) treats them as features of the model. Here they are a reliable and effective tool that do what they say on the box, but they do not license much useful inference about the structure of the actual system. They then note a second use (which they call Friston blankets) in which the Pearl blanket identified by the model is promoted to a feature of the system being modelled, so as to allow more metaphysically interesting talk. Like Andrews, they cite this as an error, but unlike Andrews, who was targeting critics of the FEP, Bruineberg et al say that FEP proponents are the ones making this error in order to make their FEP analyses more ambitious in scope.

The second critique comes from Raja et al (2021), who claim that Markov blankets are a trick that lives in the modelling strategy. They note that the Markov blanket formalism does not work on every system, that the formalism by itself does not serve as a guide to discovery of real system boundaries, and that there are important features of cognitive systems that fall outside the boundaries of what Markov blankets can do. Given these limitations, they wonder, why use Markov blankets to pick out system boundaries? They then note that the only real reason is that this mathematical formalism allows one to treat any FEP system as a Bayesian inference system, which the system being modelled may or may not be. 

These challenges place Markov blankets as features of the model, and not of the system being modelled, and point to examples in the FEP literature where this distinction is conflated in problematic ways. 

Predicting Markov Blankets

One defence of the FEP to these challenges comes from some recent work. The first paper (Hipólito et al, 2021) is a theory paper that formally predicts you should find Markov blankets in fMRI data. The second (Friston et al, 2021) is an empirical paper that does, indeed, find Markov blankets in fMRI data. Ramstead has suggested this counts as empirical evidence that Pearl and Friston blankets are, in effect, the same thing.

However, these data do not resolve the issue. The theory paper does not predict specific Markov blankets, just that there should be some in an FEP analysis of the fMRI data. Then the empirical paper also does not find specific Markov blankets in the fMRI data, just Markov blankets. That paper does not show that the blankets they found correspond to any actual neural structures (although a computational paper suggests they might be able to do so). Given the deep mathematical relationship between FEP and fMRI analyses (Friston invented both, remember), and given the reliance on computational proof-of-concept, it seems clear that these are Pearl blankets (features of the model) and there is, as yet, no evidence that they are also Friston blankets (features of the system to be modelled). Worse, I think Ramstead's argument that this refutes the above critiques just walks right in the problems they are pointing out. 

Do Markov Blankets Give Us Real Neural Parts?

Based on the evidence to date, and the presence of as yet unanswered real concerns about whether the broad inferences drawn from FEP analyses about real systems are legitimate, it seems clear to me that Markov blankets do not identify real parts and processes of the system being modelled. They are features of the model, instead; interesting tools, but not explanatory. 

Based on my previous post, I have another straight-forward reason to doubt that Markov blankets give us real parts. In that post, I argued that the grounding scale for explanations of behaviour is the scale of the organism-environment system, specifically the scale of the information that allows these two to be coupled into a single system. The FEP was developed entirely separately from considerations of the behavioural scale, and instead is about the neural scale (for the study of behaviour, anyway). But work done without reference to the grounding scale will never reveal the real parts of the system! Effectively, the FEP currently is like a scheme for decomposing an alarm clock with a hammer, instead of a screwdriver. 

Can we save the FEP? Can we start over, use the ecological analysis of a given perception-action loop, and then formalise the neural scale contribution using the formalism of the FEP? This was essentially the recommendation of earlier Bruineberg work (blogged here and here). They were looking at whether the ecological or inferential framing of the FEP was best. The fact that the FEP licenses both interpretations is evidence that is does not come with an ontology (Raja et al note this too), and Bruineberg and colleagues defend the argument that the ecological ontology is the best one for the FEP to use. At the time I noted that my take was that it was clear the FEP needed ecological psychology, it remains unclear whether the ecological approach needs the FEP. 

Overall, I remain unconvinced that the FEP will be the best way for an mechanistic ecological neuroscience to go, but I remain open to the idea that it is an option. Work to be done!