Phantastikos
Part 5: Why our brain is a predictive organ, and how these predictions are the root of all uncertainty.
The image above is called an illusory contour popularized by Gaetano Kanizsa in 1976. It contains three black pie shapes and three black, angled lines. What the image definitely does not contain is a white triangle. Yet, that is precisely what we see: a white triangle occluding a black triangle outline and three black circles. We can clearly see the contour of a shape that is not there. How can that be? Here is a related example: Have you ever walked down the stairs with your mind wandering, and misplaced the last step? Perhaps you thought yourself at the bottom when instead there was another step. Yikes, you think and tumble awkwardly to the floor. Or maybe you had arrived at the bottom, but falsely assumed there was another step. In this case, although there is no danger of falling, a strange loss of balance occurs despite standing firmly on the ground. Arms flapping and body slumping awkwardly. That is because the complex chain of movements making up a step has been interrupted mid-way. You are left in a bewildered state. You were so sure there was another step when in fact there wasn’t! And lastly, this example:
When I was a kid I could go to the store with just 5$ and come home with milk, bread, eggs, a magazine, and even some candy. These days you can’t do that any more…
…too many surveillance cameras.
The above three examples trigger a specific reaction in us. These reactions are different from one another, yet they all share the same origin. They are caused by prediction error. The first example highlights the false prediction of a geometric shape our eyes perceive although it is not explicitly there. In the second example, we encounter an error in predicting the shape and location of the ground on which to step on — with potentially catastrophic consequences. The third example illustrates a universal source of humor, a false prediction of how the story would conclude. The last line of the joke performs a surprising switch of semantic context, giving the entire story an unexpected new meaning. Why is it that we process complex visual and motor tasks so effortlessly, are easily tricked by visual illusions, and amused by unexpected language? The answer lies in the inner workings of our brain.
Since the days of Plato, sense perception was understood to be a passive process. It was thought that the input from our senses is projected onto the allegorical cave wall of our mind. This projected image was understood to be the direct representation of the world. Our eyes, Plato thought, were the apertures of the pinhole camera of our mind. This made intuitive sense to a great many eminent men with imposing beards to stroke, and culminated in the mind-body dualism defended by René Descartes. He argued that mind and matter were two distinct substrates of the world, and claimed that an immaterial soul inhabits our body, interacting with it via the pineal gland of the brain. Descarte’s statement rests on an idea humanity has been saddled with since the Bronze Age. Despite some progress in an intellectual battle that has lasted centuries, this idea is still the source of suffering and confusion to the present day. It is, in my own view, the single most pernicious, dangerous, and deadly idea ever conceived by mankind. The concept of the immortal soul. Since time immemorial, it has been violently exploited to compel, coerce, and conquer. One supernatural idea among many, it nevertheless confers great power to those who wield it by making use of an asymmetry. The asymmetry lies in the presumed benefits of an eternal afterlife on one hand, and the mundane concerns about one’s present wellbeing. A multitude of holy men have convincingly argued that the former clearly outweigh the latter, and therein lies its immense power. The immortal soul turns life into a dim passageway to be hastily traversed en route to eternal bliss. A nasty, brutish, and short transition to the land of promises. Claiming that life has no intrinsic value, it has been used to inflict untold suffering on legions of our ancestors, and we will explore it in more detail in a later chapter about the manipulation of uncertainty. Here, we will focus on the implications this idea has had on our understanding of the mind.
The belief in a soul as the source of our agency has led to the implicit assumption that a homunculus, an etherial little person, inhabits our body, observing the sense projection happening in the theater of our mind. However, in the early twenty-first century, an entirely different picture started to emerge from a growing body of neuroscientific research. Instead of a passive canvas onto which the world is being projected by our senses, our mind appears to be an advanced prediction machine. In part, this is because our brain, despite its one hundred billion neurons, would be unable to process all of the combined sensory input in real time. Researchers at Penn State have estimated that each human eye transmits ten million bits of data to the brain every second.¹ And then there are the other senses. Contrary to popular belief, we have more than the five senses of sight, touch, taste, hearing, and smell. Many more. The inner ear provides information on balance, nociceptors inside our organs detect pain, proprioceptors inform us about the orientation of our arms and legs, and osmoreceptors transmit perceptions of hunger, thirst, or nausea. In a resting state, our brain consumes about twenty percent of the body’s energy. This increases under heavy cognitive load, and would be much more if our brain had to process all sensory information in real time. So instead of trying to compute this flood of information, our brain does something creative.
The predictive brain
Our visual system is a good example of this process. It handles the complex task of turning a two dimensional patch of light into meaning. This starts within the eye, where information is spatially encoded and contours enhanced. Our retina is a computational outpost of our brain, performing “edge computing” to save bandwidth and improve response time. This compressed data stream is then sent down the optical nerve to the Lateral Geniculate Nucleus of each brain hemisphere. There, major objects are tagged with velocity of movement and relayed to the visual cortex. The signal is then separated into object, background, foreground, context, movement, saliency, et cetera. This complex task of extracting meaning from a two dimensional patch of light is achieved by a hierarchy of neural processes in which information flows up the hierarchy towards higher levels of symbolic abstraction. This sounds a little academic, so let us consider a practical example. Let’s say that your eyes detect a high-contrast outline. This information flows upward to the next hierarchical level and is recognized as a dark rectangle in front of a brown background. This signal is passed to the next level and tagged as a stationary black box on the wooden floor about one meter away from you. From there it is relayed further up and tagged as a shoe box you then recognize as a gift from your mother-in-law. In this way, your visual system can move from the detection of a raw, luminous signal, to the semantic understanding of the object observed.
Of course, this brief description is vastly oversimplified. In reality, the processes in the visual system are less linear and involve many other factors like visual attention, gaze shift, depth, self-motion, and many other parts of the brain like the pulvinar, inferior temporal cortex, hippocampus and others. At best, I have sketched a simplified flowchart of information moving up neural hierarchies to extract meaning. But there is a bigger problem with my description. In this flowchart, information only flows in one direction. Up. However, research on the visual cortex shows information pathways descending in the opposite direction.² Those descending connections even appear to outnumber the ascending connections. According to Karl Friston, a British neuroscientist at the University College London, these descending connections serve the purpose of prediction. In his view, the brain is not a passive filter of sensations, but an active inference organ that generates hypotheses that are tested against sensory evidence.³ Friston calls it a fantastic organ, from the greek phantastikos: the ability to create mental images. This means that our brain actively builds a predictive model of the world, constantly checking it against sensory input. In order to minimize surprise, our brain is organized into feedback loops of descending predictions and ascending sensory input. According to Friston, these feedback loops include dedicated channels for what he calls gain control. Gain control is a channel that applies a specific weight to each prediction. This weight indicates how confident a prediction is. A low weight represents a highly tentative guess. A speculative hunch at best. A high weight means a prediction is a highly certain one. One that is expected to be true. This means that expectations are tagged by how accurate they are believed to be. This process of weighting predictions is meta-predictive. It makes predictions about predictions. We could say that this process tags predictions by their degree of certainty. By inverting this statement we can also say that our brain uses this process to encode uncertainty. From this perspective, the element of uncertainty is the key component in the fantastic organ that is our brain. It is the meta-ingredient that makes all predictions possible. It does so by facilitating a recursive process that minimizes surprise. All of our cognitive abilities depend on it. Despite the prediction error examples at the beginning of this chapter, our brain is exceedingly good at making predictions based on partial information. Riding a bicycle downhill on difficult terrain, understanding the content of partially occluded images, or capturing the meaning of a conversation despite loud background noises are tasks we perform with ease. It only becomes apparent how computationally complex they are once we attempt to program a machine to perform them. Even the most sophisticated supercomputer can do neither of the above tasks. Our brain has these advanced abilities because of its ability to make accurate predictions based on very little information. This ability is made possible by encoding degrees of uncertainty into the system. Uncertainty is hard-wired into our neural processes. By minimizing surprise, our predictive processes deal in the currency of uncertainty.
World building machine
This is a radical departure from the traditional understanding of how the brain works. Instead of a passive canvas onto which the world is being projected, our brain is now a fantastic organ. An active world building machine. This predictive processing framework has attracted a growing number of supporters because it explains the descending connections in the organization of brain networks, but also sheds a new light on previously unexplained psychological observations. If our brain actively builds and maintains an elaborate virtual model of the world, then it makes it easier to understand any number of common observations. It would explain why people dream, why schizophrenia causes delusions indistinguishable from reality, or why psychoactive substances cause hallucinations. It also explains how our brain avoids being overloaded by the continuous tsunami of sense information. Instead of trying to compute all signals in real time, the brain builds a predictive model and updates it only when needed. A 2013 article in the Journal of Neuroscience about auditory prediction states that:
“Successively complex neural processes are involved in predicting the future, and thereby try to explain away expected prediction errors generated at lower levels. Only information about failures in prediction flows up the hierarchy, and results in the revision of expectations about the future.” ⁴
This point is key. Our understanding of the world is heavily influenced by our expectations, and beliefs shape our interpretation of facts. Our expectations descend the neural pathways of our brain, are tested against sensory input, and corrected only when they fail. Prediction failure is how the brain updates our virtual model of the world. It is how we learn. This makes intuitive sense. When walking around a house, we expect it to have a rear side. When picking up a rock, we expect it to have weight. When visiting Brian in accounting, we expect a lack of dress sense. When our expectations are wrong, we are stumped and we learn. Learning only occurs when predictions fail. And so we have come full circle back to our examples from the start of this section. Kanizsa’s triangle, the awkward tumble on the stairs, and my humble attempt at humor are all examples of failed predictions. We commonly call this prediction error by another name. Surprise. Our brain continuously builds forward models of the world and learns by dealing in the currency of surprise. By minimizing surprise, our brain builds the elaborate world of our expectations. As we have seen in the previous section, whether you are an E. coli bacterium or a human, making predictions from finite knowledge is the root of all uncertainty. But this is not just an abstract property of information processing. Uncertainty, in the shape of gain control channels, is built into the plumbing of every brain’s neural machinery. Without it, we would not be able to make accurate predictions. Without accurate predictions, our choices would frequently lead to mistakes. And making mistakes, for most life forms on this planet, is the road to extinction. Uncertainty is the shadow cast while making predictions. It follows us everywhere and affects us every time we make a choice. Like a shadow, it can cause fear. It can be made to appear larger or smaller, and it can be used to manipulate people. We cannot escape or control it, but we can chose not to be its slave. Properly understood and domesticated, uncertainty is a useful companion. This skill is a necessary ingredient in all creativity and innovation. It helps us make better decisions in unpredictable situations. Most importantly, it is foundational when dealing with the hypercomplex, dynamic environment of the twenty-first century.
Footnotes
1 Lead author Kristin Koch, in the lab of senior author Peter Sterling, PhD, Professor of Neuroscience. Study co-authors are Judith McLean and Michael A. Freed, from Penn, and Ronen Segev and Michael J. Berry III, from Princeton University. The research was supported by grants from the National Institutes of Health and the National Science Foundation.
2 van Kerkoerle T, Self MW, Dagnino B, Gariel-Mathis MA, Poort J, van der Togt C, Roelfsema PR. 2014 Alpha and gamma oscillations characterize feedback and feedforward processing in monkey visual cortex. Proc. Natl Acad. Sci. USA 111, 14 332–14 341. (doi:10.1073/pnas.1402773111)
3 Kanai R, Komura Y, Shipp S, Friston K. 2015 Cerebral hierarchies: predictive processing, precision and the pulvinar. Phil. Trans. R. Soc. B 370: 20140169. http://dx.doi.org/10.1098/rstb.2014.0169
4 Srivas Chennu, Valdas Noreika, David Gueorguiev, Alejandro Blenkmann, Silvia Kochen, Agustín Ibáñez, Adrian M. Owen and Tristan A. Bekinschtein. Journal of Neuroscience 3 July 2013, 33 (27) 11194–11205; DOI: https://doi.org/10.1523/JNEUROSCI.0114-13.2013
