The blogosphere is humming about Free Energy Principle – an interesting and general take on brain function that sees the brain as performing only one task: maintaining proper homeostasis of the organism through the good use of an internal model of the environment. What does this theory purports to be? What problems and concerns can be identified with this approach? Does it have a chance of becoming a number one theory of the mind and brain?
You can read great posts about this idea on Sergio Graziosi’s blog: The predictive brain (part one): what is this about? and The predictive brain (part two): is the idea too generic?, and on Conscious Entities: Minimising Free Energy. In this post I present you my take on this very interesting issue.
Free Energy Principle
Free Energy Principle (see also “Free Energy Principle on Wikipedia) is the idea that the brain has exactly one general function, which is minimization of free energy. This free energy is the upper bound of entropy – or disorder – of the organism. If the free energy is very high, then there is a high probability that the organism will be found in an unexpected (for the given species) state, for example “being underground” is a very unexpected state for a bird. If the entropy of an organism is high, the organism is essentially dissolving, decomposing, as it cannot maintain itself anymore. In simple terms, free energy principle states that the brain’s main function is to maintain the organism within certain bounds which allows it to survive.
An agent can minimize its free energy in two ways: by acting on the world (modifying the world for it to conform to the agent’s beliefs and models of the world) or by learning about the world (adjusting its model of the world) so that it will be better at predictive future events and generally acting in the environment (Friston & Kiebel, 2009).
Central to the free energy principle is the idea of an internal model of the world. The internal model is the knowledge about the environment and about relationships between events. Such a model is generative, that is it is used by an agent to make inferences about the world and about future events. This is closely related to what is called Bayesian brain hypothesis.
On this account actions are behaviors directed outwards that lead, in the long run, to maintaining homeostasis of the agent, to minimizing free energy, thus making sure that disorder will not overwhelm the organism.
Perception is then inference of the causes of sensory stimulation. An agent has a hypothesis about that it should see, it has a concrete model of the environment and the situation. Perception is the inference of the elements in the world that would most likely produce sensory stimulation (Friston & Kiebel, 2009).
Learning is construed as updating the model of the world based on the discrepancy between what did the model predict and what was actually perceived – the so called prediction error (Friston & Kiebel, 2009).
A critique of Free Energy Principle
Free Energy Principle (FEP) seems like a very neat and simple idea that would aid us in understanding and explaining brains and minds of various species, including our own, and help us in developing artificial intelligence. If it is so general, then it should as well be applicable to simulated minds.
However, I have a few objections and questions about the idea and its applicability to explaining psychology.
How can we understand how thoughts arise and are constructed? Clearly, if the free energy principle underlies the most basic functions of the mind, then it should tell us something new about our thoughts (imaginations). What the FEP tell us is that thoughts happen so that the organism will minimize its free energy in the long run. Which is to say that we think to survive. Not really revealing, I must admit.
As the example above shows, FEP cannot help us in understanding how specific functions of the mind work. How do we learn? How do we plan and make decisions? How do emotions interact with other thought processes? FEP won’t point us to any specific answers. We can of course formulate hypotheses, perform experiments and we do that. So far we didn’t need FEP and today it’s not clear that we do. We can of course reformulate our findings in terms of free energy, but FEP itself doesn’t naturally lead us on paths to fertile grounds for research.
The generality of FEP will not tell us what mental phenomena we can encounter in the future. Right now FEP has been applied to explaining a few cognitive processes, but only post hoc. We expect from a good theory of mind / brain predictions about what we will find. If the FEP won’t give us that in the next decade, then we will be much more inclined to say that FEP acts only as a good description tool that we can use to characterize what we already know.
Central to the FEP account of the brain is the idea of an internal model that is used to make inferences and to which sensory experiences are compared. What exactly is an internal model?
Similarly, how can we define ideas and concepts from the FEP perspective and how can we explain their interactions?
FEP states that hypothesis about an external cause of a sensory experience, if found to be wrong, is used to update the agent’s internal model. How does this updating happen? This is a known optimization problem, with which AI researchers are struggling for some time. Does FEP give us any pointers how can this work? Error back-propagation is a technique used in machine learning in algorithms that teach neural networks how to behave better and better. In this technique we attempt to modify the network in such a way as to minimize its errors. What do we need FEP then? Sure, we can say that these algorithms minimize free energy of the network, but I’m not sure if any progress has been made due to the FEP itself.
An agent is thought of as producing inferences, or hypotheses, about the environment and future events. How are these hypotheses formulated in the first place?
What does FEP has to say about language and emotions (except, of course, that they minimize free energy)? Is it useful in understanding, researching, explaining language?
Free Energy Principle as a general theory of the mind
From a good general theory of the mind and brain we would expect a capacity to formulate new questions and methods of researching the phenomena in question, to unify known knowledge in an understandable framework, to tell us something new about minds and brains that we didn’t know before (to facilitate making predictions about mental phenomena), generally to tell us how to understand and how to study minds and brains.
There already are some theory (of varying generality) that try to unify explanations into a coherent framework. A few of the examples of such theories are: Computationalism, embodied cognition (Barsalou, 2010; Shapiro, 2012), dynamicism (Hotton & Yoshimi, 2011; Spivey, 2007).
Free Energy Principle is posited as a general theory of the mind (Friston, 2010). I applaud all the efforts in making the theory better and better. I have already mentioned my reservations for it above. I must conclude that – at least for the time being – FEP doesn’t give us enough tools, conceptual apparatus and a vision as to posit it as the most fruitful and the most prospective general theory of the mind.
I myself observe the rapid progress in making new discoveries, in developing new methods, new conceptual perspectives in disciplines of complexity theory, dynamical systems theory, network theory. I hope that these endeavors will neatly fit with research in neuroscience and generally in cognitive sciences, leading to a valuable and general theory of the brain and the mind. That’s why I myself am a proponent of a Complex Mind Theory, that I see as a promising perspective on studying the mind as a complex dynamical system of a special sort.
Bayesian approaches to brain function. (2014, April 9). In Wikipedia, The Free Encyclopedia. Retrieved 12:37, February 11, 2015, from http://en.wikipedia.org/w/index.php?title=Bayesian_approaches_to_brain_function&oldid=603449548
Computational theory of mind. (2014, December 26). In Wikipedia, The Free Encyclopedia. Retrieved 14:55, February 15, 2015, from http://en.wikipedia.org/w/index.php?title=Computational_theory_of_mind&oldid=639664169
Embodied cognition. (2015, January 28). In Wikipedia, The Free Encyclopedia. Retrieved 14:59, February 15, 2015, from http://en.wikipedia.org/w/index.php?title=Embodied_cognition&oldid=644572856
Free energy principle. (2015, January 31). In Wikipedia, The Free Encyclopedia. Retrieved 12:17, February 11, 2015, from http://en.wikipedia.org/w/index.php?title=Free_energy_principle&oldid=644977855
Barsalou, L. (2010). Grounded cognition: past, present, and future. Topics in Cognitive Science, 2(4):716–724.
Friston, K. (2010). The free-energy principle: a unified brain theory?. Nature Reviews Neuroscience, 11(2), 127-138.
Hinton, G. E. (2007). Learning multiple layers of representation. Trends in Cognitive Sciences, 11 (10), 428-434. 10.1016/j.tics.2007.09.004
Hohwy, J. (2015). The Neural Organ Explains the Mind. In T. Metzinger & J. M. Windt (Eds). Open MIND: 19(T). Frankfurt am Main:MIND Group. Doi: 10.15502/9783958570016
Hotton, S. and Yoshimi, J. (2011), Extending Dynamical Systems Theory to Model Embodied Cognition. Cognitive Science, 35: 444–479. doi: 10.1111/j.1551-6709.2010.01151.x
Shapiro, L. (2012). What’s new about embodied cognition? Filosofia Unisinos, 13(2-suppl.):214–224.
Spivey, M. (2007). The continuity of mind, volume 40. Oxford University Press.
Clark, A. (2013). Perceiving as Predicting. 9th International Symposium of Cognition, Logic and Communication “Perception and Concepts” at University of Edinburgh, UK. https://www.youtube.com/watch?v=05P41FQlgjI
Friston, K. (2014). Consciousness and the Bayesian brain. Sandler Conference. https://www.youtube.com/watch?v=HeQfO4byFhg