Free Energy, Creativity, Neuroscience, and Psychoanalysis

Author: Prof. Jim Hopkins

This is a handout from Prof Hopkins' presentation at the MPTT Winter Conference on 13th of January 2017.

Introduction

An increasing number of neuroscientists are indicating interest in the paradigm developed under the free energy principle by Karl Friston and his colleagues.

This approach, according to which the brain operates to minimize free energy, seems particularly important for psychiatry. It has been advocated in two articles in Lancet Psychiatry for 2014: Corlet and Fletcher’s ‘Computational Psychiatry: a Rosetta Stone Linking the Brain to Mental Illness’ and Friston et al ‘Computational Psychiatry: the Brain as a Phantastic Organ.’ It is also particularly important – and I think genuinely a Rosetta Stone -- for medical psychotherapy, and for a reason not discussed in those articles.

Friston’s account of the working of the brain in minimizing free energy partly replicates that proposed in Freud’s Project for a Scientific Psychology, and hence the overall role of free energy in Freud’s account of the mind. This replication is extensive enough to make it reasonable to suppose that psychoanalytic theory can be fairly directly integrated into neuroscience, using the free energy approach in conjunction with the affective neuroscience recently developed by Jaak Panksepp, Antonio Damasio, and others. This suggests a radical expansion in the possibilities for collaboration with other approaches, as well as for confirmation and disconfirmation.

We will consider some aspects of neuroscience and creativity integration below, but several effects seem likely. First, insofar as medical psychotherapists would adopt a common model of the working of the brain, disputes as to the effectiveness of different kinds of therapeutic intervention would become easier to deal with.

A second consequence might be a more unified view of the underlying causes of mental disorder. In a psychoanalytic perspective there is a clear sense in which a great range o disorders – schizophrenia, bipolar disorder, depression, mania, OCD, and others – have an underlying common cause. They stem from basic emotional conflicts to which everyone is liable. This is not reflected in symptom based classification schemes like those of DSM, which notably yield many co-morbidities involving supposedly distinct disorders. Like psychoanalysis, much recent work in neuroscience – e.g. that concerning cortical thinning across psychiatric classifications, and in areas that seem involved in emotional conflict – seems consistent with an underlying common causes.

Finally psychoanalytic thinking would also naturally be expanded. Consider, for example, the emerging neuroscientific consensus that REM dreaming is a final phase in the consolidation of memory, and one where the emotions linked with old memories are reactivated. This would let us see why dreams might have the kinds of dense and deep meanings that analysts find in them, and would also prompt further research combining psychoanalytic and neuroscientific claims on the connections between dreams, memories, and emotions.

Primary and Secondary Processes

  • Like phantasy, the primary process shielded the infant from the barrage of external (exteroceptive) and internal (interoceptive and proprioceptive) sensory impingements that follow birth. The most important of these were the ‘endogenous (interoceptive) stimuli stemming from the ‘peremptory demands’ of bodily needs (and other biological imperatives). These aroused the aversive feelings and emotions connected with ’the trauma of birth and the infantile anxiety of longing’ for the mother. The primary process countered these aversive emotions by producing wish fulfilling phantasies of the satisfaction of the imperatives involved in accord with what Freud called the pleasure principle.

  • In addition the primary process enabled the infant to learn, by paving the way for the secondary processes of realistic perception, memory, thought and purposive action. In this it worked in accord with the reality principle, enabling the infant to take a progressively more active role in securing biological imperatives in co-operation with the mother and others. Insofar as this was successful it led to the satisfaction of the imperatives and the pacification (inhibition) of the aversive impingements and emotions that failure to satisfy them aroused. Thus the secondary processes displaced the primary process by actually satisfying biological imperatives in waking activity, and thereby relegated pure primary process functioning to the production of the fictive experiences of dreams as the guardians of sleep.

Thus from Freud’s early neuroscientific presentation the dialectic between the primary and secondary processes (like that between phantasy and thought) determined a theory of the working of the mind and brain that encompassed both development and disorder.

Development partly consisted in using the reality-oriented secondary processes to secure biological imperatives, thereby pacifying the aversive impingements/arousals stemming from their frustration. But since this depended on success in co-operating with the mother and others, it was liable to disruption by failure in this task -- which might stem either from internal or external sources.

Such failure entailed a vicious circle of blighted development of the secondary processes in the co-operative securing of imperatives and excessive dependence on the primary process (phantasy) in responding to the aversive impingements/arousals generated by their frustration. This included the daytime incursions of the primary process that characterize the various mental disorders.

Prediction and Thinking as Hypothesis Formation

One distinctive alteration in the core theory concerned the notions of expectation (or tacit prediction – expecting something to be a certain way) and hypothesis.

Hanna Segal argued that from birth the infant was engaged in ‘testing phantasy against reality’ and in this came prepared with ‘expectations’ in the form of ‘unconscious phantasy’. She regarded these expectations as a form of hypothesis, which could be altered by thinking if they were not borne out in experience. Thus she regarded the infantile phantasy that good experiences were caused by a breast/mother and bad experiences by a bad breast/mother as a kind of hypothesis that was modifiable by thought in light of experience.

In this the displacing of the primary process by the secondary process as hypothesized by Freud was seen as a process in which innate phantasy gave way to perception - or experience-based thought, and thought was a process in which prior hypotheses were replaced by others that better explained the course of experience, and because the tacit predictions (expectations) of the prior hypotheses were not borne out. Such an account of thinking -- as the tacit formation and confirmation of hypotheses that explain the course of experience – is a very plausible one.

Brain as Operating to Minimize Free Energy

In the work of Friston and his colleagues, as in the speculative neuroscience of Freud’s Project, the brain is conceived as operating to minimize a quantifiable free energy. In both this free energy is introduced by sensory impingement with ‘endogenous [interoceptive] impingements being a particularly important source. In Freud these impingements reflect ’the major needs’ as in Friston they reflect compliance with biological imperatives, predicting departures from an overall (homeostatic or allostatic) equilibrium (Pelluzio at al, 2015).

As Freud speaks of these imperatives as creating a ‘demand on the mind for work’ to produce ‘specific actions’, Friston (2012) speaks of ‘an imperative to minimize [free energy]’ by producing the ‘kinematic trajectories’ of bodily movement, including purposive action. For both this requires the brain to embody a model of the world, including the agent’s body (in Freud the ‘bodily ego’) and in the elaboration of Freud by Segal above this requirement is initially met by the innate generation of a prior or phantasy model of reality, whose generation is replaced by that of progressively more realistic models over the course of development, so that the generation of the virtual or phantasy version of reality is relegated to dreaming.

Virtual Reality Generator

Thus in the most recent treatment of these topics, Hobson, Hong, and Friston (2014) cite evidence that during the third trimester of pregnancy infants spend most of their time in REM, and argue that ‘the brain is genetically endowed with a virtual reality generator’ whose working ‘is most clearly revealed in [REM] dreaming’.

They thus hold that ‘we are born with a virtual reality model’ of what we will subsequently discover to be the causes of internal and external sensory impingement. As well as innate this model is intrinsically predictive, and so is ‘entrained by sensory prediction errors’, particularly during the sensory initialization attending birth, to become ‘a generative or predictive model of the world’, whose functioning in virtual reality mode is, as in Freud, relegated to REM dreaming.

Hence just as in Freud the primary process version of the world is displaced by that based on perception, memory, and action, so in Friston an initially virtual version of the world is displaced by one based on perception, memory, and action. Freud’s secondary processes are realized in Friston by active inference, which involves the active collection of sensory data integrated with purposive action by which free energy is minimized.

These structural parallels make clear that Friston’s version of free energy neuroscience admits of the same systematic linkage between development and disorder as we saw in Freud.

Just as development in Freud partly consists in using the secondary processes in waking life to replace the primary process and the phantasy version of the world that it generates, so development in Friston partly consists in using active inference in waking life to replace the operation of the virtual reality generator and virtual version of the world that it generates. Just as in Freud this replacement takes place partly via the success of the secondary processes in securing biological imperatives, so in Friston it takes partly via the success of active inference in the same process.

It follows that Friston’s account admits the possibility that Freud took as his principal hypothesis, namely that the replacement of the primary process (virtual reality generator) in waking life might be inadequate or incomplete, leaving the individual subject to the generation of forms of virtual reality constitutive of mental disorder. Friston has welcomed this connection and has encouraged others to pursue it, and his collaborator on dreams Alan Hobson, for many decades a particularly vehement critic of Freud, has espoused it writing that his current work ‘takes up the Project for a Scientific Psychology exactly where Freud left it in 1895’ (Hobson 2015 p. 4) To consider this we must take up Friston’s account in more detail.

Friston’s Notion of Free Energy as Derived from Helmholtz’s Perceptual Neuroscience

Freud’s speculative conception of free energy seems to have been derived from Helmholtz’s account of free energy as the energy in a system that was available for conversion into work. Friston’s originated in another claim by Helmholtz, namely that we should understand vision – and perceptual processes generally – as involving the tacit generation of hypotheses about the objects that are the causes of perceptual experience, via which hypotheses the brain operates to predict the impingements on its sensory surfaces that produce these experiences.

In his view, similar in some respects to that of Segal scouted above, the brain embodies a model of the world, which it uses to predict sensory impingements, and to generate the conscious perceptual experiences of objects that these impingements cause. Thus as he said each movement we make by which we alter the appearance of objects should be thought of as an experiment designed to test whether we have understood correctly the invariant relations of the phenomena before us, that is, their existence in definite spatial relations.

Such a model as Helmholtz envisaged would have to be a statistical model, based on the relative frequencies of differing types of sensory impingement.

We can see this particularly clearly in the case of vision, where we readily recognize how fleeting and perishing the actual impingements registered by our sensory receptors are.

Such fragmentary but numerous impingements could only produce stable conscious perceptual experiences of their causes if the brain subjected them to some form of statistical analysis that linked perceptual representations of the causes to very large numbers of the impingements (and this conclusion is reinforced by considering the multimodal nature of perception.)

This is illustrated by the data about saccadic eye movements depicted on the next three slides, which also illustrate what Friston means by active inference in the case of vision.

Experiments on prediction of visual sensory input when reading

  • Purposive conceptual guidance of saccadic fixation.
  • Saccadic Fixation and Perceptual Acuity in Reading
  • Purposive Conceptual Guidance of Saccadic Fixation in Reading

The act of reading a paragraph thus illustrates an ongoing process of active and predictive inference on the part of the brain which is both conceptually guided and hierarchical.

Overall the process is governed by the reader’s ongoing grasp of the meaning the text, which depends on her hypotheses about the author’s intentions in writing it.

This in turn guides the reader’s hypotheses about the individual words, which are continuously confirmed or disconfirmed by 1/5 of a second saccadic sampling of a few letters of each word, with each hypothesis about the word currently under the fixation yielding, via the incorporation of the data from this sampling, the hypothesis that yields the next saccadic fixation point.

All this of course goes on below the level of consciousness, but at the same time the reader is checking her conscious hypotheses about the author’s communicative intentions in writing, via the registration of the unfolding meaning of the sentence being read.

This, however arises from her continual unconscious recomputation of the predicted meanings via her grasp of her hypotheses as to the meanings of the words and the way they are combined in the sentence.

The hypothetical nature of this whole hierarchical process is well illustrated by experiments in which eye-tracking machines are used to anticipate the reader’s saccades, and so to fool the reader about the text she is reading by changing everything except the letters that are the focus of the current saccade. This is described by Daniel Dennett.

When your eyes dart about in saccades the muscular contractions that cause the eyeballs to rotate are ballistic actions whose trajectories at lift-off will determine where they will hit ground zero at a new target. If you are reading text on a computer screen your eyes will leap along a few words with each saccade, farther and faster the better a reader you are. A computer equipped with an eye-tracker can detect and analyse the lift-off in the first milliseconds of a saccade, calculate where ground zero will be, and, before the saccade is over, erase the word on the screen at ground zero and replace it with a different word of the same length. What do you see? Just the new word, and with no sense at all of something having been changed. As you peruse the text on the screen, it seems to you for all the world as if the words were carved in marble, but to another person reading the same text over your shoulder (and saccading to a different rhythm) the screen is aquiver with changes.

The effect is overpowering. When I first encountered an eye-tracker experiment, and saw how oblivious the subjects were (apparently) to the changes flickering on the screen, I asked if I could be a subject. I wanted to see for myself. While I waited for the experiment to start I read the text on the screen. ’Why don’t you turn it on?’ I asked. ‘It is on’ they replied.

Conscious experiences as hypotheses predicting and explaining the data of sensory impingement

All this indicates that we can best understand the brain as Helmholtz suggested, that is, as embodying a statistical model whose hypotheses or parameters are precise and powerful enough both to predict the data of sensory impingement from hypotheses about the causes of the impingements, and also to predict the causes from the data of impingement.

Can brain build a model of the world?

This, however, takes us to a second problem, which is familiar from philosophical arguments promoting scepticism about the senses. How can the brain, lodged within the skull, construct a model of things that are in the environment outside the skull? Clearly it does so, and, as Helmholtz speculated, by predicting the data of impingement. But how does this guarantee that the model is actually correct? For as we just saw, the model can definitely be wrong.

In the 1990’s Geoffrey Hinton (and colleagues) realized that the answer to this was to think of the brain as minimizing information-theoretic free energy, a construct that the celebrated physicist Richard Feynman had derived to assess models in statistical physics. Free energy in this sense is a measure of the evidential support given to a Bayesian statistical model of some data by the capacity of the model to explain and predict the data. Strikingly its mathematical formulation is also traceable to Helmholtz, and is derived from Helmholtz’s mathematical formulations defining physical free energy (hence the Helmholtz-related terminology of computational neuroscience).

Friston’s Notion of Free Energy as Derived from Helmholtz’s Perceptual Neuroscience via Richard Feynman and Geoffrey Hinton

Feynman proved that the free energy of a model defines a bound on the evidential support that the model enjoys from the data it predicts and explains. The greater the free energy, the lesser the evidential support. So decreasing the free energy is a way of increasing this evidential support.

What makes free energy the answer to the problems we have been considering is the fact that it can be assessed and minimized by using features of the model itself – by calculation involving just the parameters or hypotheses that compose the model, and the success or failure of these in predicting the data that the model is supposed to predict and explain.

Thus, as Hinton realized, if the brain did embody a model of the causes of the impingements on its sensory receptors, it would be certain to improve itself as a model by altering its own hypotheses or parameters from within, in such a way as to minimize its own errors in the prediction of impingements. In perceptual learning it would do this by altering the hypotheses that generate erroneous predictions so as to generate better ones.

The brain would do this, moreover, in a hierarchical way that corresponds the levels upwards from saccadic fixation through word recognition through word meaning through sentence meaning through the understanding of the author’s intention to communicate in the example of reading above.

As we saw, the process of reading was driven by top-down predictions from the higher levels of this hierarchy, and so could proceed with relatively little contact with the data, with predictions being checked and altered only via the 1/5 second saccadic skips. In this process the data of impingement serve as perceptual input only insofar as they are not as predicted, in which case the unpredicted datum proceeds up the neurocomputational hierarchy until it causes an alteration in a hypothesis or parameter that would have enabled that parameter to predict it, and stops there as the model moves on.

Hence the sensory data only need to be processed when the sensory impingements are not as predicted.

What is free-energy?

Free-energy is basically prediction error. From our emotions we recognize prediction errors as surprise. Small errors mean low surprise.

In the 1990’s Geoffrey Hinton and colleagues developed such an approach in detail, by thinking of the brain as a ‘Helmholtz machine’. This was a multi-layer artificial neural network that built a statistical model of the causes of the input to its own ‘sensory’ layer. These consisted of artificial neurons, storing information via the strength of the ‘synaptic’ connections between them, and the model was hierarchical, in that each layer in the multi-layer network learned to predict activity in the next layer down. This was not representing the brain as a digital computer, rather it was creating a computational device whose working was modelled on that of the brain.

But this was also an hypothesis about how the brain itself worked. Applying it in detail to the brain, Friston realized that the same idea could be used to explicate motor behaviour as well as perception, and so could be applied from perceptual input through motor output to every aspect of its working.

Statistical models of the kind embodied (and learnt from data) by such a network enable the network to move between ’sensory’ inputs to the network and the causes of these inputs external to the network. So given a particular set of inputs, the model can predict the most probable cause, and given a particular cause, the model can predict the most probable inputs.

The notion of (variational) free energy now used in computational neuroscience was originally used to characterize these networks. Free energy in this sense is a measure of the predictive competence of any predictive statistical model of this kind – that is, any model that characterizes the causes of the data it is devised to explain precisely enough to allow the probabilistic prediction of the data from the causes and the causes from data. The notion was originally developed by the celebrated physicist Richard Feynman as a measure of the predictive competence of statistical models in theoretical physics, and although it does not describe physical energy, is it mathematically isomorphic with Helmholtz’s notion of free energy in physics.

The reason it is so important is that it is effectively calculable on the basis of data available within the model itself.

Appendix 1: Eye tracking dependence on context

References

-- Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 36(3), 181–204.

-- Corlett, P. R., & Fletcher, P. C. (2014). Computational psychiatry: a Rosetta Stone linking the brain to mental illness. The Lancet Psychiatry, 1(5), 399–402.

-- Damasio, A., & Carvalho, G. B. (2013). The nature of feelings: evolutionary and neurobiological origins. Nature Reviews Neuroscience, 14(2), 143–152.

--Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138.

-- Friston, K. J., Stephan, K. E., Montague, R., & Dolan, R. J. (2014). Computational psychiatry: the brain as a phantastic organ. The Lancet Psychiatry, 1(2), 148–158.

-- Genzel, L., Spoormaker, V. I., Konrad, B. N., & Dresler, M. (2015). The role of rapid eye movement sleep for amygdala-related memory processing. Neurobiology of Learning and Memory, 122, 110–121.

-- Hallett, P. E., & Lightstone, A. D. (1976). Saccadic eye movements towards stimuli triggered by prior saccades. Vision Research, 16(1), 99–106.

-- Hobson, J. A., Charles, C.-H. H., & Friston, K. J. (n.d.). Virtual reality and consciousness inference in dreaming. Retrieved from https://www.researchgate.net/profile/KarlFriston/publication/267743554Virtualrealityandconsciousnessinferenceindreaming/links/568bfedf08ae153299b634f7.pdf

-- Hopkins, J. (2016). Free Energy and Virtual Reality in Neuroscience and Psychoanalysis: A Complexity Theory of Dreaming and Mental Disorder. Frontiers in Psychology, 7. https://doi.org/10.3389/fpsyg.2016.00922

-- Panksepp, J. (2011). Cross-species affective neuroscience decoding of the primal affective experiences of humans and related animals. PloS One, 6(9), e21236.

-- Panksepp, J., & Biven, L. (2012). The Archaeology of Mind: Neuroevolutionary Origins of Human Emotions. W. W. Norton & Company.

-- Panksepp, J., & Solms, M. (2012). What is neuropsychoanalysis? Clinically relevant studies of the minded brain. Trends in Cognitive Sciences, 16(1), 6. Underwood, G. (2005). Cognitive Processes in Eye Guidance. OUP Oxford.

-- Yarbus, A. L. (1967). Eye Movements During Fixation on Stationary Objects. In Eye Movements and Vision (pp. 103–127). Springer US http://link.springer.com/chapter/10.1007/978-1-4899-5379-7_4