07/18/2020
I’ve recently been more excited about a scientific idea than I’ve ever seen before, and that’s predictive processing. The general idea is relatively simple, but the implementation seems really complex. It seems to be a holistic theory of what the brain is doing, at a level of complexity that seems about right. Predictive processing seems to be right in the middle of what neurons are doing and what the temporal lobe is for in terms of complexity, which gives us room to move up and down those layers in this framework. It’s also fundamentally computational in nature, which is attractive to me because I have a bias for the idea that everything is computation.
The basic idea of predictive processing is this: what we perceive is a computational reconciliation between a “top-down” stream of information and a “bottom-up” stream of sensory information. This is what we’re going to attempt to unpack.
Let’s start with the bottom-up stream. This stream is comprised of both our outer senses (i.e. vision, hearing, etc.) and our interoceptive senses, which just means information about our body’s internal state. This can include a sense of hunger, info about our own blood sugar, a sense of where our limbs are in relation to the rest of the body, etc. So our brains are constantly getting multiple “input” streams, from a variety of sources, at once.
The top-down stream of information is made up of concepts encoded in our prior experience. This stream can be thought of as a constant stream of predictions about what the bottom-up sensory stream is going to look like next.
Prediction Error Minimization
Where the streams meet, a computation occurs. According to predictive processing, a comparison is made between what was predicted by the top-down stream and the actual sensory information.
Mismatches between predictions and actual sensory input are not used passively to form percepts, but only to inform updates of representations which have already been created (thereby anticipating, to the extent possible, incoming sensory signals). The goal of these updates is to minimize the prediction error resulting from the prediction (feature #5, Prediction Error Minimization (PEM)), in such a way that updates conform to the norms of Bayesian Inference [1]
Alright, hopefully this isn’t too overwhelming. If it helps, think about the top-down stream of information as your expectations, and when the top-down can’t match the bottom-up, the experience is one of confusion. I hesitate to use such imprecise words as “confusion” and “expectation,” but I think it’s useful here. Imagine you take a drink of vodka expecting it to be water, the brief experience of confusion there is intense and unpleasant, and your knowledge of the world is very quickly updated as you spit it out.
These error minimization computations happen hierarchically. At every level, the two streams meet to try to resolve any error. In the case of water-but-expecting-vodka, at the lowest level, a prediction of the taste of water streams in from above. When the prediction is compared with the actual sensory signal, that level says to the one above, “We got it way off. See what you can do.” So that level tries to resolve the error, gets it way off, and levels of the hierarchy keep relaying it to different levels to attempt to resolve the error. I think this is a great segue into how action is incorporated into predictive processing.
Action
Fundamentally, there are only two ways an organism can resolve prediction error. One is internally, by continually searching for a prediction that better fits the sensory signals. The other is by changing the sensory signals to fit the prediction. As agents capable of movement, we are able to alter our sensory signals through action. In the previous example, this is equivalent to resolving error by spitting out the vodka.
Here, “action” doesn’t have to mean interaction with the external world. We mentioned interoceptive signals earlier, and these can be altered by “action” in interesting ways, too. Let’s say your brain gets word that your blood sugar is low. Your perception is a craving for sugary things, and your brain has two options to resolve the error: 1) eat a sugary thing (interaction w/ the external world) or 2) metabolize fat stores (interaction w/ internal bodily functions) [1].
This has an interesting implication:
In short, the error between sensory signals and predictions of sensory signals (derived from internal estimates) can be minimized by changing internal estimates and by changing sensory signals (through action). What this suggests is that the same internal representations which become active in perception can also be deployed to enable action. This means that there is not only a common data-format, but also that at least some of the representations that underpin perception are numerically identical with representations that underpin action.
This is called the Ideomotor Principle. In this model, action and perception have the same neural representation. I don’t fully understand this idea yet, but I don’t think it’s necessary to dive into for this introduction.
Explanatory
I am attracted to this theory because it seems to have such explanatory power, at least regarding many things related to our conscious experience. Attention, for example, is explained as “the process of optimizing precision estimates.” [1] Makes sense. We become pretty certain pretty quickly about whatever our attention is on. There’s also explanations for cognitive diseases like schizophrenia (too much trust in top-down prediction) and autism (overwhelming flows of bottom-up information). Dreaming can be explained as the interaction of concepts in your body of top-down knowledge in the absence of bottom-up information to correct error with. Placebo results have a good explanation through this lens: if we expect less pain, our brain will smooth the noisy pain signals, assuming they are partly a mistake.
Models of the world are only as good as they are explanatory/predictive. Predictive processing is the most accurate model of brain computation I’ve ever learned about, so I’ll be diving deeper into it and sharing more of what I find. Descriptions of the brain in terms of action potentials or brain lobes have left me unsatisfied. When trying to answer the question, “What is the brain doing?” a computational model seems most apt.
If you want to know more, definitely check out the source I pulled most of this info from below. And if this stuff fascinates you like it does me, check out the book Surfing Uncertainty by Andy Clark.
[1] https://predictive-mind.net/papers/vanilla-pp-for-philosophers-a-primer-on-predictive-processing