Can the Free Energy Principle be made ecological? (Bruineberg et al, 2018)

Everyone loves Friston's free energy principle (FEP), and everyone wants it for their own. Not everyone can have it, though (well, at least not if it's going to mean anything) and so there's a spirited fight about who's theory it best fits in the literature. 

Bruineberg, Kiverstein & Rietveld (2018) argue two points in an effort to win the fight for the good guys. First, they want to show that inferential, representational takes on the FEP end up in an unworkable place. Second, they want to show that an ecological/enactivist analysis works much better. Overall I think they take a solid swing at both, so it will be interesting to see the responses this sparks. Here I want to review their arguments.

To unbury the lede, I like this paper a lot. It's really long and repetitive, but in here is an excellent ecological analysis of the free-energy principle that also works to explicitly rule out the competition. I am obviously biased, but their work pointing out the flaws of Hohwy's account all make good sense to me, not least because these flaws show up in all kinds of places in the representational ontology. Hohwy fails for the reason interface theory does, in my view, and it's nice to see separate analyses end up in the same place as me. 

For what it's worth, I am not yet convinced that the FEP is the way we need to go. However, if it ends up being a good idea, Bruineberg et al have done sterling work in showing how we should go about it. 

The Free Energy Principle

The free energy principle (FEP) begins with the hypothesis that the defining features of living systems is that they actively work to maintain their states and forms. To do this, the organism must find the right subset of states to be in that it is able to maintain. These are relatively high-probability states, because otherwise the organism would be dead. Mathematically, these states have a low surprisal. States other than the optimal ones are less probable, and so if an organism finds itself in one of those states, surprisal goes up. Self-maintenance, therefore, is the business of minimising surprisal. 

Organisms can't detect surprisal, but they can (in principle) detect a quantity called free energy, and this turns outs to be an upper bound on surprisal. Minimising free energy must therefore reduce surprisal, and the hypothesis is that this is the basic mechanism at work for living and cognitive systems as they work to keep themselves intact and functioning. 

There are two ways to reduce free energy, which is (roughly) a measure of how attuned the internal and external dynamics are (or not). The first is (roughly) perception ('perceptual inference') - this changes the internal dynamics of the organism towards the demands of the external. The second is (roughly) action ('active inference') - this changes the external dynamics towards the demands of the internal. 

Bruineberg et al then review two ways of fleshing this basic idea out so as to apply to cognition, theirs and the more inferential, Helmholtzian approaches. I'll do the latter first. 

1. Helmholtz can't make the FEP work right

Bruineberg et al look to the work of Hohwy (2013) as the exemplar inferential take on the FEP, the prediction-error minimisation framework. 

In this approach, free energy minimisation in cognition is a brain-based operation. Hohwy takes the brain to be living in a skull-shaped vat, with no access to anything other that it's own states (spike trains, that sort of thing) and a flow of sensory data. The brain predicts sensory states, and surprisal goes up when that prediction is not doing so well. Free energy minimisation is implemented by reducing the prediction error, either by changing the brain's hypothesis that makes the predictions, or by acting by experimenting on the world to bring it in line with the predictions. 

Bruineberg et al don't think this works. Hohwy treats perception and action (perceptual vs active inference) as separate processes that work in opposite directions, and that each are available as options. Earlier in the paper, however, Bruineberg et al show (pg 2424) that perceptual inference cannot affect surprisal without active inference, and highlight that this is important. For example, if a human found themselves in a too-hot shower but simply altered their internal dynamics to accommodate this fact, they would burn. These cannot be separate processes - perceptual and active 'inference' must work together to get the organism into a better state. 

...on a Helmholtzian account of the mind, the aim of perceptual inference can be said to infer the most likely cause of sensory input, the more objective the better. However, we would argue that a perfect hypothesis that precisely represents the state of the environment is worthless if it does not specify what action minimises surprisal...' (pg 2430)

2. Affordance-based perception-action implements the FEP

The strong positive claim in this paper is that an ecological/enactivist analysis is the only one that can fully implement the obligations of the FEP. (I'll note here that this of course includes these authors commitment to affordances being relations, which they aren't. But I don't think that causes any problems here, so I'll leave my bitching about this to this paper.)

Briefly, surprisal becomes a measure of the disattunement of internal and external dynamics. As the organism dynamics swing in and out of attunement with the environmental dynamics, they shift in their sensitivity to the landscape of affordances. For example, say my blood sugar dips and I move into the less-than-optimal state of 'hungry'. Perception/perceptual inference is implemented by me altering my internal dynamics, not to accomodate to the hunger, but to become selectively sensitive to food affordances (my grip on the field of affordances changes). This then immediately makes me more likely to act/do active inference that will, in fact, reduce free energy and therefore surprisal. When hungry, I will move so as to eat food because those affordances are the ones dominating my behaviour. I therefore end up back in more probable, more functional states. 

This implements the full perception-action free-energy-minimising loop, and does so at an organism (rather than just brain) level. Instead of active inference figuring out which brain hypothesis is true about the world, it instead leads the organism to perceive what needs to be done to maintain itself (pg 2432).

3. Other things

They also spend some time defending the idea that, while the internal and external dynamics are separable, the brain is not in any sense in a vat. They do this by noting that the idea of Markov blankets creates a natural reason to talk about 'inside' and 'outside' of things while still allowing these things to be meaningfully coupled. Based on this, they also point out that there is nothing compulsory about the term 'inference' that Friston uses, and (like with the Watts governor) that gloss adds nothing useful to the explanation. 

References

Bruineberg, J., Kiverstein, J., & Rietveld, E. (2018). The anticipating brain is not a scientist: the free-energy principle from an ecological-enactive perspective. Synthese, 195(6), 2417-2444.

Hohwy, J. (2013). The predictive mind. Oxford University Press.