This isn't a book for the general reader! Andy Clark is a professor of Logic and Metaphysics at Edinburgh University and a lot of this book deals with the latest ideas from Artificial Intelligence. So he is quite capable of writing "Computationally frugal solutions stressing embodiment, action, and the exploitation of bodily and environmental opportunities emerge quite naturally from a predictive processing (PP) framework involving cascading inference, internal generative models, and ongoing estimations of our own uncertainty." (p 112) and that is by no means the most difficult sentence you will encounter. On the other hand, there are moments when phrases such as "The profound cognitive entanglement of brain, body, and world" (p 83) make this book sound almost erotic.
But what really hit me was the brilliance of the ideas. And, geek that I am, ideas of this elegance which are capable of bearing as much fruit as these ideas can are ideas that I find mind-blowingly sexy.
For centuries, philosophers have alternated between the empiricist idea that all we know is what we learn from our senses, which pictures the brain as a fundamentally passive lump of jelly being bombarded by sensory data, and the solipsistic ideas that all we know is our thoughts and if there is a world out there is is one created by our fertile imaginations. It's, like, all in the mind, man.
The trouble is that there is powerful evidence for the latter point of view. For example, context dependent optical illusions where we interpret an ambiguous mark as, for example, a B or a 13 depending on whether the context has led us to expect a letter of a number. It seems clear that our expectations are shaping our perceptions. Seeing may be believing but, in cases like this, believing is seeing.One third of those told to listen to a degraded recording of White Christmas and to signal when the song started 'recognised' the onset of the song even though the tape contained nothing but white noise.
So Clark suggests that we do indeed think by sending a stream of predictions out into the world. These shape what we perceive and, indeed, direct our attention to seek evidence to confirm those predictions. But the sense data that then comes back in is (to some extent) different from our expectations so we generate an error message. We use this error message to refine our expectations.Clark suggests that generating predictions and then modifying them through error detection is a "computationally frugal" process which is used by at least one form of data compression. If we want to know what happens next when viewing a running man, rather than recompute every pixel we can simply move the entire image two pixels to the right, pick up the anomalies, and compute these few anomalous pixels.
We know we do this. Our brains 'fill in' the blind spot in our visual field by making the assumption that the images around it are continued within it. Although our eyes move in lots of little jerks as we scan a scene, because our brain knows that our eyes are moving, and how they are moving, we make the appropriate adjustments and see the scene as stationary.We can use these expectations (which he calls hyperpriors when used globally) to fool ourselves. When we hear speech and see lips move we assume that the two are linked; this is how we are fooled by ventriloquists. "It is surprisingly easy ... to induce ... the illusion that a rubber hand, placed on the table in front of you, is your own. The illusion is created by ensuring the the subject can see someone tapping the realistic rubber hand, while (just out of sight) their own hand is being tapped in exact synchrony." (p 197)
Of course, the fact that we use expectations to shape and to select perceptions can lead to vicious circles. If Jill expects Jack to be angry she will selectively notice behaviours that reinforce her hypothesis, increasing her expectations. (p 73)But, excitingly, this model of the brain as being an active system, always interacting with the world, explains:
- why we notice things that aren't there: "if we hear a regular series of beats and then a beat is omitted, we are perceptually aware (quite vividly aware) of its absence. Moreover, there is a familiar sensation of 'almost experiencing' the onset of the omitted item." (p 89)
- dreaming: "During sleep, precise prediction errors are not generated [because there is no sense data], so the balance shifts towards the reduction of model complexity. Sleep may thus allow the brain to engage in synaptic pruning so as to improve (make more powerful and generalizable) the knowledge enshrined in the generative model." (p 101)
- memory: "We are built to act in ways that are sensitive to the contingencies of the past, and that actively bring forth the futures that we need and desire." (p 111)
- why you can't tickle yourself: "the feeling of ticklishness requires a certain element of surprise" (p 113) and therefore the generation of expectation prior to the performance of the action prevent such "self-predicted sensations" (p 114)
- The role of 'mirror neurons' in empathising with other people: By deploying predictions "we may sometimes grasp the intentions of other agents" (p 139) by treating them as "context-nuanced versions of ourselves" (p 139). "Human infants, around the age of 4, possess not only a sense of themselves as individual agents with specific needs, wants, and beliefs but also a sense of others as distinct agents with their own needs, wants and beliefs. How might this be achieved? The discovery of 'mirror neurons' has seemed, to many, to deliver a substantial part of the answer." (p 151)
- Placebo-based pain reduction
- Schizophrenia: "Delusions and hallucinations ... might flow from ... falsely generated and highly weighted (high-precision) waves of prediction error ... it is the weighting (precision) assigned to these error signals that makes them so functionally potent, positioning them to drive the system into plasticity and learning, forming and recruiting increasingly bizarre hypotheses .. such as telepathy and alien control." (p 206) "Once such higher level stories take hold, now, low-level sensory stimulation may be interpreted falsely. When these new priors dominate, we may thus experience hallucinations that appear to confirm or consolidate them." (p 207) Schizophrenia may this stem from false error signals. This may in turn explain why they are less able to smoothly track with their eyes a moving object which is "one of the most widely replicated behavioral defects in schizophrenia" (pp 208 - 209 and why they are better able than neurotypical subjects to self-tickle (p 212).
- Autism: "Autistic subjects are less susceptible to illusions in which prior knowledge is used to interpret ambiguous sensory information" (p 224). If prior knowledge is attenuated ie weaker hyperpriors ie less influence of context, subjects are likely to "treat more incoming stuff as signal and less as noise" (p 225) which in turn will mean that "huge amounts of incoming information are treated as salient and worthy of attention, thus increasing effortful processing" (p 225) which is likely to "contribute to the emergence of a variety of self-protective strategies involving repetition, insulation, and narrowing of focus" (p 225). "The social domain is highly complex ... in which context ... is everything and in which the meaning of small verbal and non-verbal signs must be interpreted against a rich backdrop of prior knowledge" (p 225) which might explain why autistic subjects often have poor social skills.
Yes, it is sometimes hard wading through the prose, but when you hit nuggets such as these you know there is gold in these here hills!
August 2016, 300 pages