51动漫

ON the 300th anniversary of Johannes Kepler鈥檚 death, Albert Einstein said: 鈥淚t seems that the human mind has first to construct forms independently before we can find them in things.鈥� He was referring to Kepler鈥檚 astounding deduction that the orbits of planets around the sun were not circular as scholars had believed, but elliptical 鈥� a feat that would set the stage for Newton鈥檚 laws of motion.

Were Einstein alive today, he might be amazed at his own prescience. Modern neuroscience seems to agree with him, as it struggles to explain how 1400 grams of gelatinous stuff in a bony skull creates an inner world. 鈥淭he mystery is, and remains, how mere matter manages to give rise to thinking, imagining, dreaming, and the whole smorgasbord of mentality, emotion, and intelligent action,鈥� writes Andy Clark in his new book, Surfing Uncertainty. 鈥淏ut there is an emerging clue.鈥�

The clue is prediction. As Einstein surmised, our brains do indeed seem first to construct forms in order to find them in the world outside. Our bodies are constantly being bombarded by sensations. Our brains have to make sense of that chaotic, often uncertain input, and they do so every moment by making predictions, or educated guesses, about what might be generating the signals impinging on us. How that leads to perception, action and an embodied mind comprises the 鈥減redictive processing story鈥�, synthesised in delicious detail in Clark鈥檚 book.

The traditional bottom-up view of visual perception, for example, holds that our brain analyses incoming signals, finds patterns of ever-increasing complexity, and makes sense of what鈥檚 out there by matching observed patterns against internal representations. Predictive processing turns these notions upside down.

In this paradigm, which has its roots in ideas developed by German physician Hermann von Helmholtz in 1860, our brains actually generate sensory data to match what鈥檚 coming in, using internal models of the world and of our bodies. These 鈥済enerative models鈥� give rise to multiple hypotheses about the sources of the incoming sensory data, and the most likely hypothesis becomes a perception.

But this is an ongoing process. The brain compares generated with incoming data, identifies any errors and updates its internal models as necessary, so that it can predict and thus perceive more accurately the next time around.

The paradigm, says Clark, is one of a multilayered brain, where predictions flow down from higher cortical layers to progressively lower layers, and error signals flow from the bottom up as well as laterally within each layer. Errors can be given more or less credence, depending on the context. It鈥檚 all about minimising uncertainty.

This way of thinking about the brain hides something profound: in order to generate sensory data, the brain must know something about the way the world works and how bodies move in this world. But no one is feeding it textbooks. So when the brain creates accurate internal models 鈥� over evolutionary timescales or an organism鈥檚 lifetime 鈥� it is akin to understanding the physical world. Indeed, it is akin to understanding ourselves as embodied minds navigating our environments.

鈥淭o make sense of chaotic inputs, the brain makes educated guesses as to what generates them鈥�

Surfing Uncertainty is full of examples that are strongly suggestive, bordering tantalisingly on proof, that this is how brains work. One that swayed me involves sine-wave speech 鈥� a degraded, coarse version of a speech recording. You can try to make sense of it yourself by finding an online example.

Listen hard: it鈥檒l be pure R2-D2. Next, listen to the original sentence, then to the degraded version. This time, you will start to hear the words, albeit patchily. That鈥檚 because your brain now has prior knowledge, which it uses to generate hypotheses about what鈥檚 creating the sine-wave speech. The best hypothesis is what you perceive. It鈥檚 a striking example of the brain making meaning where none existed.

The fact that we are active agents ties neatly into this predictive processing story. Brains have to reduce prediction errors if we are to perceive the world correctly 鈥� a must for survival. They can, of course, do this by 鈥渇inding the predictions that best accommodate the current sensory inputs鈥�, explains Clark.

They can also do this by making our bodies move, forcing them to seek sensations so that hypotheses can be verified or rejected, refining internal models. As Clark writes, neatly alluding to the book鈥檚 title: 鈥淲e are not cognitive couch potatoes idly awaiting the next 鈥榠nput鈥�, so much as proactive predictavores 鈥� nature鈥檚 own guessing machines forever trying to stay one step ahead by surfing the incoming waves of sensory stimulation鈥�.

This is a book for which most of us will need to brace ourselves: it is more monograph than popular science, which means the going is tough. But grappling with technical language and concepts is repaid with fresh insights and intensely evocative moments.

For example, Clark sees the predictive brain as an 鈥渁ction-oriented engagement machine鈥�, constantly configured and reconfigured by the body and the environment, which in turn are being shaped by a voracious brain as it seeks to make sense of the outside world. As he puts it: 鈥淭his pattern repeats at more extended scales of space and time, as we structure (and repeatedly restructure) the social and material worlds that slowly but surely structure us.鈥�

鈥淲e are not cognitive couch potatoes idly awaiting the next 鈥榠nput鈥�, so much as proactive predictavores鈥�

Cognitive processes like attention also fit well into the story. Attention, in predictive processing, assigns high integrity to incoming data, so that any prediction errors are taken more seriously and internal models updated as needed. On the other hand, if you are doing something mostly without paying conscious attention, prediction errors don鈥檛 carry the same import, and internal models hold sway.

Crucially, the fact that a brain filled with such internal models can generate sensory data on its own, independent of actual sensations, allows us to imagine, to dream and to time-travel mentally. We owe our inner lives to 鈥減robabilistic generative models鈥�, as obtuse as that sounds. This inner life includes perception of our own body: according to the predictive-processing idea, even our sense of being embodied emerges from generative models that try to infer the causes of both external and internal sensations (such as signals from the gut).

But predictive processing, it turns out, comes with a dark side. Predictions can go awry, leading to delusions and hallucinations, for example, in people with schizophrenia. This also shows how perception and cognition may not be as distinct as we once believed: what we know and how we think can influence perception, which in turn modulates thought and knowledge.

Of course, predictive processing can鈥檛 explain how consciousness emerges from the material brain. But Clark is optimistic. 鈥淢ight we, inch-by-inch and phenomenon-by-phenomenon, begin to solve the so-called 鈥榟ard problem鈥� of conscious experience?鈥� he asks. 鈥淚t is far too early to say, but it feels like progress,鈥� he answers.

In Surfing Uncertainty, Clark makes a formidable case for predictive processing, one that synthesises the work of many well-known researchers, such as Geoffrey Hinton and Karl Friston. But he is suitably cautious in not overstating the case, peppering his book with caveats for balance: 鈥淚f the predictive processing story is on track鈥� is one favourite.

There are dangers in being seduced by the allure of predictive processing. But proof, one way or the other, may come from our efforts to create intelligent and conscious agents, and from seeing how they do what we do.

It may be a tall order, but as Nobel laureate Gerald Edelman and Giulio Tononi pointed out in their book A Universe of Consciousness: 鈥淎fter all, it has been done at least once by evolution.鈥� They meant our brain, of course.

This article appeared in print under the headline 鈥淧redicting the world鈥�