A real wealth of papers at the OpenMind site, presided over by Thomas Metzinger, including new stuff from Dan Dennett, Ned Block, Paul Churchland, Alva Noë, Andy Clark and many others. Call me perverse, but the one that attracted my attention first is the paper The Neural Organ Explains the Mind by Jakob Hohwy. This expounds the free energy theory put forward by Karl Friston.
The hypothesis here is that we should view the brain as an organ of the body in the same way as we regard the heart or the liver. Those other organs have a distinctive function – in the case of the heart, it pumps blood – and what we need to do is recognise what the brain does. The suggestion is that it minimises free energy; that honestly doesn’t mean much to me, but apparently another way of putting it is to say that the brain’s function is to keep the organism within a limited set of states. If the organism is a fish, the brain aims to keep it in the right kind of water, keeps it fed and with adequate oxygen, and so on.
It’s always good to get back to some commonsensical. pragmatic view, and the paper shows that this is a fertile and flexible hypothesis, which yields explanations for various kinds of behaviour. There seem to me to be three prima facie objections. First, this isn’t really a function akin to the heart pumping blood; at best it’s a high level meta-function. The heart does blood pumping, the lungs do respiration, the gut does digestion; and the brain apparently keeps the organism in conditions where blood can go on being pumped, there is still breathable air available, food to be digested, and so on. In fact it oversees every other function and in completely novel circumstances it suddenly acquires new functions: if we go hang-gliding, the brain learns to keep us flying straight and level, not something it ever had to do in the earlier history of the human race. Now of course, if we confront the gut with a substance it never experienced before, it will probably deal with it one way or another; but it will only deploy the chemical functions it always had; it won’t learn new ones. There’s a protean quality about the brain that eludes simple comparisons with other organs.
A second problem is that the hypothesis suggests the brain is all about keeping the organism in states where it is comfortable, whereas the human brain at least seems to be able to take into account future contingencies and make long-term plans which enable us to abandon the ideal environment of our beds each morning and go out into the cold and rain. There is a theoretical answer to this problem which seems to involve us being able to perceive things across space and time; probably right, but that seems like a whole new function rather then something that drops out naturally from minimising free energy; I may not have understood this bit correctly. It seems that when we move our hand, it may happen because we have, in contradiction of the evidence, adopted the belief that our hand is already moving; this belief serves to minimise free energy and our belief that the hand is moving causes the actual movement we believe in.
Third and worse, the brain often seems to impel us to do things that are risky, uncomfortable, and damaging, and not necessarily in pursuit of keeping our states in line with comfort, even in the long term. Why do we go hang-gliding, why do we take drugs, climb mountains, enter a monastery or convent? I know there are plenty of answers in terms of self-interest, but it’s much less clear to me that there are answers in terms of minimising free energy.
That’s all very negative, but actually the whole idea overall strikes me as at least a novel and interesting perspective. Hohwy draws some parallels with the theory of evolution; like Darwin’s idea, this is a very general theory with alarmingly large claims, and critics may well say that it’s either over-ambitious or that in the end it explains too much too easily; that it is, ultimately, unfalsifiable.
I wouldn’t go that far; it seems to me that there are a lot of potential issues, but that the theory is very adaptable and productive in potentially useful ways. It might well be a valuable perspective. I’m less sure that it answers the questions we’re really bothered about. Take the analogy of the gut (as the theory encourages us to do). What is the gut’s function? Actually, we could define it several ways (it deals with food, it makes poop, it helps store energy). One might be that the gut keeps the bloodstream in good condition so far as nutrients are concerned, just as the lungs keep it in good condition in respect of oxygenation. But the gut also, as part of that, does digestion, a complex and fascinating subject which is well worth study in itself. Now it might be that the brain does indeed minimise free energy, and that might be a legitimate field of study; but perhaps in doing so it also supports consciousness, a separate issue which like digestion is well worthy of study in itself.
We might not be looking at final answers, then – to be fair, we’ve only scratched the surface of what seems to be a remarkably fecund hypothesis – but even if we’re not, a strange new idea has got to be welcome.
The Open MIND collection is cutting edge, to be sure! I highly recommend Hohwy’s book, The Predictive Mind, as well.
There’s a lot the approach doesn’t explain, but it really is remarkable how many otherwise baffling phenomena it does explain (the book does a good job laying out his case). Most importantly it solves the cognitive bootstrapping problem, the question of how perception can even get off the ground given the Inverse Problem (the problem of inferring causes from effects).
I agree that analogizing the brain’s functions to organ function generally is problematic. Regarding your third worry, a big part of Friston’s picture involves ‘active inference,’ the way the systems are designed to use behaviour to secure the information required to minimize free energy. The ‘comfort’ or ‘risk’ encountered at the organism level is actually neither here nor there – thus the saying, ‘Curiosity killed the cat…’ 😉 We often do stupid things in the name of error-minimization.
Personally, I think it’s unfortunate that Hohwy seems intent on using Friston’s account to rationalize more classical (and historically intractable) cognitive science commitments. The mechanics/semantics dichotomy plagues his account as profoundly as it does any other. The difference is that Hohwy *actually has the mechanisms he can specify.* This is as good an example of mechanical complexities producing ‘semantic effects’ as you could ask for. So the question for Hohwy, then, is why he nevertheless insists that the semantics comes first, that these mechanisms are performing honest to goodness ‘inferences’ over and above what they do mechanically? Why bother with ‘spooky functionalism’ once you have a naturalized account? Clark, I think, is more honest about the way this work, far from securing traditional verities, raises a ‘desert of the real’ spectre, providing, for instance, ways to talk about behaviour absent ‘goals.’
Yeah, I don’t know about the whole free energy thing, but the idea of the brain as just one more interdependent organ is, uhh, thought provoking, insofar as it does then present the opportunity to explore in that context just what that organ’s function is. And what it is not.
While I intuitively agree, I don’t necessarily think you’ve made the case that the brain has the higher order function. I mean, the heart keeps the organism in conditions wherein the others can do their thing, as well. In fact, very few of our organism’s organs provide results that we could do without. Seems more symbiotic than hierarchical to me.
Of course, I also understand that you know a lot more than I do about this, and that I may be missing the background to have an informed opinion.
Not at all – that’s a fair point.
Peter, now I’m egocentrically convinced that you’re trying to troll me…
Hohwy’s book is in my living room, waiting for its turn (it’s the next in line), and I haven’t read the article you link here, but I have given this whole area a good deal of thought (and of course a bit of study/reading to start with).
Your first objection: understanding the brain as another organ. Yes, the brain is an awkward fit, but what does this observation tell us? To me, as usual, it exposes our cognitive limitations: in order to believe that we understand a part of our body, we give it a name and associate it with a function. Lungs are for breathing. Which is of course oversimplifying and shifting the problem at the same time (what is breathing? what is breathing for?). So, this objection speaks about how we think, making it relevant here, but in a very indirect way… I’m surprised that Scott didn’t jump on this one.
Second and third objections are very relevant indeed, and are amongst the reasons why the general consensus is that you need a lot of legwork to build upon the FEP and arrive to a complete explanation of brains (let alone minds). I would add, incidentally, that we are talking about very abstract, mathematical principles, and that one could and should expect that their embodiment would be somewhat scrappy and inelegant, because of physical constraints, evolutionary pathways and in general, the messiness of reality. There are very practical implications when people start doing neuroscience with FEP as a guiding principle, but perhaps it’s best to leave this discussion aside.
I have a fourth objection, though, which leads to “embodied cognition” and the like. FEP tells us that the whole organism tends to minimise free-energy, not the brain. It also tells us that: 1. modelling the external environment is necessary to effectively minimise free-energy. 2. therefore active inference follows (you already have a model, using it to make predictions is an easy extension that produces better results).
My objection is: nothing of the above mentions brains. In fact, FEP specifically applies to a single bacterium as well. We can hypothesise that the brain “is for” making active inferences, but I personally regard also this as an over-simplification. The gut will do some active inferences on its own (at a low level!), as other organs will, while some inferences will be done in collaboration with the brain, and the brain will also be doing other stuff that can’t be easily described as “producing active inferences”. Therefore we need to be cautious: expecting to explain every function of the brain (in isolation) in terms of active inferences is probably not going to work.
Un-falsifiability: in its more general and abstract version, FEP looks unfalsifiable to me. That’s the beauty of it: I see it as a meta-theory that can be used to generate falsifiable hypotheses. In my view, neuroscience has a gigantic chicken and egg problem: lots of data available, but very little in terms of theories that may be used to tie all data-interpretations together. So, to build such a theory, we need to learn more, and thus collect more data, which remains only locally-interpreted, because we lack general theories, etc. Therefore, FEP, active inference, Bayesian Brain and more “neural” hypotheses from the same family are very promising: we know that FEP should be applicable to describe/model the function of brain+body and that, in terms of computations, much of the “function” will happen in the brain. So the quest becomes one of tying the highly abstract theory on top with the raw data at the bottom: there is a direction to follow and the chicken and egg problem is solved. (TBH, the chicken and egg problem is thus scattered in a gazillion smaller problems, but that’s another story.)
I’ll wrap up. I find it really fascinating, that much is clear. Would FEP or a related concept solve the hard problem on its own? I very much doubt so, but that’s because the hard problem requires an intellectual solution (on top of empirical explanations). Yes I know, people will disagree, etc. It’s only my hunch.
On the practical side however, FEP and family look really promising. It’s a pithy that it is so notoriously hard to actually understand it.
I’ve written two posts on the subject, coming from a different angle, and trying to make the subject accessible, feedback is always welcome.
The predictive brain (part one): what is this about?
and
The predictive brain (part two): is the idea too generic?
(that’s why I’m feeling trolled, in a good way! 😉 )
This Free Energy concept reminds me of something Searle once said – that materialist theories aren’t really very good but no one wants to be the guy who opens the door for theists to come back in and stink up the place.
Pingback: Free Energy Principle – A quick look and some concerns | Observing Ideas
I see Hohwy is Deputy Editor for this interesting new venture;
http://nc.oxfordjournals.org
The journal is sponsored by the ASSC, which formerly sponsored Psyche; let’s hope the new one has a longer run.
To keep it simple, the problem I see is that the brain is considered as a whole one single organ. This applies to the heart, the lungs or the pancreas, etc that even though their cells can fulfil different functions, e.g. chemical or mechanical, and/or endocrinologic, there is a main function easy to identify. But the brain is many things in one. Even more, different parts of the brain can be in conflict and oppose each other. Maybe the argument makes sense for some functions of the brain, but not for the whole. Thus, the approach proposed in the essay is just too simplistic to be taken seriously.
On the other hand minimising free energy leads to absolute death, no more free energy, entropy at maximum levels, and nothing left to be done. I think the brain (healthy ones) strives for the opposite, to keep life moving on, no matter how, and it plays nasty tricks on us to accomplish this goal.
Amazingly, living organisms reduce their own entropy (well yes, at the expenses of increasing the surroundings’ one, after all the second principle only requires the production of entropy is positive)
I have the impression that physic concepts are borrowed too carelessly for biological and psichological matters discussions.
Machine learning researcher here, interested to see whether the Free Energy Principle will catch on as an organizing principle for “sentient robotics”. I am not aware of any other theories which hold out the promise of coherent top down designs of machines capable of handling perception, action and attention in an integrated way.
Engineers familiar with Bayesian learning would have little difficulty taking the FEP on board, although they would probably formulate it in a way which abstracts away from predictive coding as this term is used in the neuroscience literature (other forward-backward architectures such as the variational autoencoder could serve instead).
To the extent that I can understand Friston’s papers, he seems to be able to tell a plausible story in situations where fairly low level prior beliefs (concerning the causes of sensory data and goals to be achieved on various time scales) are treated as given. But has anybody got any ideas about the form of the uber-prior from which these low level priors would need to be derived in order to tell a complete story? (I am referring to the brain’s virtual reality model, whose parameters are supposedly being estimated online through (Hebbian) free energy minimization.)
I know this question calls for speculation but an engineer hoping to use the FEP would actaully have to supply some sort of answer to it. Friston’s attempt to identify the uber-prior with the Freudian ego is a testament to the glorious fertility of his imagination but it is not fit for purpose.
https://www.ncbi.nlm.nih.gov/pubmed/20194141
Thanks, Patrick, nicely put.
Thanks, Peter. I have just discovered your site and your book. Looks like I am in for a treat. Lovely writing.
@Patrick #9 & 11,
Yes, CE is a delightful place to explore – I wish I could spend more time in here!
From an engineering perspective, I’d steer away from FEP and look closer to more detailed implementations, and then abstract away useful concepts (if any). Predictive Coding is one of the numerous alternatives…
Reason is that FEP itself (as I understand it) is a theoretical lens that can be applied to reverse engineer pre-existing biological systems (i.e. we “know” they must be minimising FEP, so we can start looking at “how”, FEP provides a solid attack strategy). The solidity of this approach comes from the tautological side of FEP, see the “part 2” link in my comment #4.
I have been exploring this subject for quite some time in my own time (and blog), covered an engineering perspective (in very general terms) in at least one post, now; might give you some ideas, if I did it right. More lucubrations will follow, when I’ll find the time. I’d be interested in your feedback if you’ll find it interesting.
I am pleased to meet you, Sergio. Your blog was the first resource I turned to after coming across Friston’s Aeon article. I then read Andy Clark’s book but I found that I still couldn’t understand what Friston was talking about. So I took the plunge and started reading Friston’s papers …
I will follow up on your blog.
Thanks Patrick! I was hoping to put together something useful, hope it somewhat worked for you… Subject is tricky and Friston isn’t really good at making his ideas approachable (understandably) ;-). The format of “journal article” doesn’t help either.
Pingback: A short response to Karl Friston’s Free Energy theory – Your Mileage May Vary