The “Bayesian” brain as a “hierarchical prediction machine” is an enticing new perspective on old problems, for all the reasons Clark articulates, ranging over fields as disparate as neuroanatomy, artificial intelligence, psychiatry, and philosophy; but he also catalogues some large questions that need good answers. While waiting for the details to come in, I want to suggest some other benefits that this perspective promises. If it turns out not to be sound, in spite of all the converging evidence Clark describes, we will have all the more reason for regret.
It is everybody's job – but particularly the philosophers' job – to negotiate the chasm between what Wilfrid Sellars (Reference Sellars and Colodny1962) called the manifest image and the scientific image. The manifest image is the everyday world of folk psychology, furnished with people and their experiences of all the middle-sized things that matter. The scientific image is the world of quarks, atoms, and molecules, but also (in this context particularly) sub-personal neural structures with particular roles to play in guiding a living body safely through life. The two images do not readily fall into registration, as everybody knows, leaving lots of room for confusion and compensatory adjustment (nicely exemplified by the surprise/surprisal pair).
Consider what I will call Hume's Strange Inversion (cf. Dennett Reference Dennett2009). One of the things in our world is causation, and we think we see causation because the causation in the world directly causes us to see it – the same way round things in daylight cause us to see round things, and tigers in moonlight cause us to see tigers. When we see the thrown ball causing the window to break, the causation itself is somehow perceptible “out there.” Not so, says Hume. This is a special case of the mind's “great propensity to spread itself on external objects” (Treatise of Human Nature, Hume Reference Hume and Selby-Biggs1739/1888/1964, I, p. xiv). In fact, he insisted, what we do is misinterpret an inner “feeling,” an anticipation, as an external property. The “customary transition” in our minds is the source of our sense of causation, a quality of “perceptions, not of objects,” but we mis-attribute it to the objects, a sort of benign user-illusion, to speak anachronistically. As Hume notes, “the contrary notion is so riveted in the mind” (p. 167) that it is hard to dislodge. It survives to this day in the typically unexamined assumption that all perceptual representations must be flowing inbound from outside.
Here are a few other folk convictions that need Strange Inversions: sweetness is an “intrinsic” property of sugar and honey, which causes us to like them; observed intrinsic sexiness is what causes our lust; it was the funniness out there in the joke that caused us to laugh (Hurley et al. Reference Hurley, Dennett and Adams2011). There is no more familiar and appealing verb than “project” to describe this effect, but of course everybody knows it is only metaphorical; colors aren't literally projected (as if from a slide projector) out onto the front surfaces of (colorless) objects, any more than the idea of causation is somehow beamed out onto the point of impact between the billiard balls. If we use the shorthand term “projection” to try to talk, metaphorically, about the mismatch between manifest and scientific image here, what is the true long story? What is literally going on in the scientific image? A large part of the answer emerges, I propose, from the predictive coding perspective.
Every organism, whether a bacterium or a member of Homo sapiens, has a set of things in the world that matter to it and which it (therefore) needs to discriminate and anticipate as best it can. Call this the ontology of the organism, or the organism's Umwelt (von Uexküll Reference von Uexküll and Schiller1934/1957). This does not yet have anything to do with consciousness but is rather an “engineering” concept, like the ontology of a bank of elevators in a skyscraper: all the kinds of things and situations the elevators need to distinguish and deal with. An animal's Umwelt consists in the first place of affordances (Gibson Reference Gibson1979), things to eat or mate with, openings to walk through or look out of, holes to hide in, things to stand on, and so forth. We may suppose that the Umwelt of a starfish or worm or daisy is more like the ontology of the elevator than like our manifest image. What's the difference? What makes our manifest image manifest (to us)?
Here is where Bayesian expectations could play an iterated role: Our ontology (in the elevator sense) does a close-to-optimal job of representing the things in the world that matter to the behavior our brains have to control. Hierarchical Bayesian predictions accomplish this, generating affordances galore: We expect solid objects to have backs that will come into view as we walk around them, doors to open, stairs to afford climbing, cups to hold liquid, and so forth. But among the things in our Umwelt that matter to our well-being are ourselves! We ought to have good Bayesian expectations about what we will do next, what we will think next, and what we will expect next! And we do. Here's an example:
Think of the cuteness of babies. It is not, of course, an “intrinsic” property of babies, though it seems to be. What you “project” out onto the baby is in fact your manifold of “felt” dispositions to cuddle, protect, nurture, kiss, coo over, . . . that little cutie-pie. It's not just that when your cuteness detector (based on facial proportions, etc.) fires, you have urges to nurture and protect; you expect to have those very urges, and that manifold of expectations just is the “projection” onto the baby of the property of cuteness. When we expect to see a baby in the crib, we also expect to “find it cute” – that is, we expect to expect to feel the urge to cuddle it and so forth. When our expectations are fulfilled, the absence of prediction error signals is interpreted as confirmation that, indeed, the thing in the world we are interacting with has the properties we expected it to have. Cuteness as a property passes the Bayesian test for being an objective structural part of the world we live in, and that is all that needs to happen. Any further “projection” process would be redundant. What is special about properties like sweetness and cuteness is that their perception depends on particularities of the nervous systems that have evolved to make much of them. The same is of course also true of colors. This is what is left of Locke's (and Boyle's) distinction between primary and secondary qualities.
The “Bayesian” brain as a “hierarchical prediction machine” is an enticing new perspective on old problems, for all the reasons Clark articulates, ranging over fields as disparate as neuroanatomy, artificial intelligence, psychiatry, and philosophy; but he also catalogues some large questions that need good answers. While waiting for the details to come in, I want to suggest some other benefits that this perspective promises. If it turns out not to be sound, in spite of all the converging evidence Clark describes, we will have all the more reason for regret.
It is everybody's job – but particularly the philosophers' job – to negotiate the chasm between what Wilfrid Sellars (Reference Sellars and Colodny1962) called the manifest image and the scientific image. The manifest image is the everyday world of folk psychology, furnished with people and their experiences of all the middle-sized things that matter. The scientific image is the world of quarks, atoms, and molecules, but also (in this context particularly) sub-personal neural structures with particular roles to play in guiding a living body safely through life. The two images do not readily fall into registration, as everybody knows, leaving lots of room for confusion and compensatory adjustment (nicely exemplified by the surprise/surprisal pair).
Consider what I will call Hume's Strange Inversion (cf. Dennett Reference Dennett2009). One of the things in our world is causation, and we think we see causation because the causation in the world directly causes us to see it – the same way round things in daylight cause us to see round things, and tigers in moonlight cause us to see tigers. When we see the thrown ball causing the window to break, the causation itself is somehow perceptible “out there.” Not so, says Hume. This is a special case of the mind's “great propensity to spread itself on external objects” (Treatise of Human Nature, Hume Reference Hume and Selby-Biggs1739/1888/1964, I, p. xiv). In fact, he insisted, what we do is misinterpret an inner “feeling,” an anticipation, as an external property. The “customary transition” in our minds is the source of our sense of causation, a quality of “perceptions, not of objects,” but we mis-attribute it to the objects, a sort of benign user-illusion, to speak anachronistically. As Hume notes, “the contrary notion is so riveted in the mind” (p. 167) that it is hard to dislodge. It survives to this day in the typically unexamined assumption that all perceptual representations must be flowing inbound from outside.
Here are a few other folk convictions that need Strange Inversions: sweetness is an “intrinsic” property of sugar and honey, which causes us to like them; observed intrinsic sexiness is what causes our lust; it was the funniness out there in the joke that caused us to laugh (Hurley et al. Reference Hurley, Dennett and Adams2011). There is no more familiar and appealing verb than “project” to describe this effect, but of course everybody knows it is only metaphorical; colors aren't literally projected (as if from a slide projector) out onto the front surfaces of (colorless) objects, any more than the idea of causation is somehow beamed out onto the point of impact between the billiard balls. If we use the shorthand term “projection” to try to talk, metaphorically, about the mismatch between manifest and scientific image here, what is the true long story? What is literally going on in the scientific image? A large part of the answer emerges, I propose, from the predictive coding perspective.
Every organism, whether a bacterium or a member of Homo sapiens, has a set of things in the world that matter to it and which it (therefore) needs to discriminate and anticipate as best it can. Call this the ontology of the organism, or the organism's Umwelt (von Uexküll Reference von Uexküll and Schiller1934/1957). This does not yet have anything to do with consciousness but is rather an “engineering” concept, like the ontology of a bank of elevators in a skyscraper: all the kinds of things and situations the elevators need to distinguish and deal with. An animal's Umwelt consists in the first place of affordances (Gibson Reference Gibson1979), things to eat or mate with, openings to walk through or look out of, holes to hide in, things to stand on, and so forth. We may suppose that the Umwelt of a starfish or worm or daisy is more like the ontology of the elevator than like our manifest image. What's the difference? What makes our manifest image manifest (to us)?
Here is where Bayesian expectations could play an iterated role: Our ontology (in the elevator sense) does a close-to-optimal job of representing the things in the world that matter to the behavior our brains have to control. Hierarchical Bayesian predictions accomplish this, generating affordances galore: We expect solid objects to have backs that will come into view as we walk around them, doors to open, stairs to afford climbing, cups to hold liquid, and so forth. But among the things in our Umwelt that matter to our well-being are ourselves! We ought to have good Bayesian expectations about what we will do next, what we will think next, and what we will expect next! And we do. Here's an example:
Think of the cuteness of babies. It is not, of course, an “intrinsic” property of babies, though it seems to be. What you “project” out onto the baby is in fact your manifold of “felt” dispositions to cuddle, protect, nurture, kiss, coo over, . . . that little cutie-pie. It's not just that when your cuteness detector (based on facial proportions, etc.) fires, you have urges to nurture and protect; you expect to have those very urges, and that manifold of expectations just is the “projection” onto the baby of the property of cuteness. When we expect to see a baby in the crib, we also expect to “find it cute” – that is, we expect to expect to feel the urge to cuddle it and so forth. When our expectations are fulfilled, the absence of prediction error signals is interpreted as confirmation that, indeed, the thing in the world we are interacting with has the properties we expected it to have. Cuteness as a property passes the Bayesian test for being an objective structural part of the world we live in, and that is all that needs to happen. Any further “projection” process would be redundant. What is special about properties like sweetness and cuteness is that their perception depends on particularities of the nervous systems that have evolved to make much of them. The same is of course also true of colors. This is what is left of Locke's (and Boyle's) distinction between primary and secondary qualities.