The Idea
What happens in our brains when we interact with the world around us? Let’s say when we see an apple in front of our eyes.
In the traditional view, light from the apple produces various sensations in our eyes. Our eyes then send these sensations to different layers of neurons within our brain. Each layer identifies different features of these sensations – colour, shape, movement, etc. Finally, all these individual features are integrated into one singular experience of “seeing an apple”. This is called perception.
Predictive Processing (PP) flips this model. It reimagines our brain as a prediction machine, always predicting its own future. But its future is influenced by various sensations produced in our sense organs when we interact with reality. So, to predict its own future, it must also predict how the outside world will interact with it.
Every time it generates these predictions, it sends them to all the relevant neurons connected to our sense organs. And every time we interact with reality, these neurons check whether the predictions they receive match the sensations produced by reality.
If they match, these neurons stay silent and our brain preserves its beliefs about external reality. These beliefs are called priors. If they don’t match, these neurons send an error signal to the neurons above them in the hierarchy. This is called prediction error. Our brain uses these prediction errors to update its priors and checks if the new priors result in accurate predictions. This process goes on over and over again until all the prediction errors become small enough to be ignored.
This is the view presented by Predictive Processing.
Why is it Revolutionary?
If you compare the two views I shared, you’ll notice something missing in the second picture: perception.
You’ll see that in this new view, our brain can’t perceive any sensations produced by our interactions with reality. Because when our predictions are right, these sensations are not even passed on to other neurons in our brains. And when they are incorrect, only the resulting prediction error is passed on.
So, most regions of our brain only receive two kinds of information:
- Predictions generated by neurons above them in the hierarchy
- Prediction errors sent by neurons below them
So our perception can only come from one of these.
Predictive Processing argues that all we ever perceive in life – colours, tones, scents – are the predictions generated by our own brains and NOT sensations produced by the outside world.
It’s like living in a Matrix-style simulation generated by our own brains. A simulation whose purpose is not to give us an accurate picture of our reality, but to correctly predict what sensory inputs are generated when we interact with it.
This is a revolutionary idea with, as we shall see later, far-reaching consequences. But first, how do we know it’s the right idea?
The Clues
Well, we don’t. But several clues are pointing us in this direction. Let’s look at a few interesting ones.
#1 The mystery of our brain’s energy efficiency
You might have heard of AI systems consuming enormous computing power and energy as they are beginning to get smarter. Yet, we all have a brain doing far more complex tasks with just a fraction of the energy.
How is our brain pulling this off? What makes it SO energy efficient?
In 2021, a team of researchers decided to investigate this question by making a small change to a class of neural networks called Recurrent Neural Networks (RNN).
In our brain, the firing of neurons consumes the most energy. So, to minimize energy consumption, we’d have to minimize the amount of firing. In RNNs (and in all neural networks), the amount of firing is influenced by the strength of the connection between its neurons – also called weights. So the researchers forced the RNN to perform its task by simultaneously trying to achieve the smallest possible weights between its neurons.
With just this one extra constraint, the network started organising itself into an architecture where some of its neurons began predicting the signals received by other neurons, which then learnt to fire only if these predictions were wrong. Initially, when the RNN had no data to make accurate predictions, it generated a lot of prediction errors, resulting in a lot of firing. However, as its predictions improved, the firing rates decreased, minimising the energy consumed.
This is a pretty significant clue – forcing regular neural networks to conserve energy naturally results in a predictive processing architecture.
#2 A brain slice that learnt to play a videogame
In a fascinating experiment from 2022, now popular as “Dishbrain”, researchers connected a bunch of neurons grown in a petri dish to a simple video game – Pong, where you control a paddle to hit a ball back and forth.
One part of the brain tissue was fed signals about the ball’s position and the signals from a different part of the tissue were used to move the paddle. In Pong, every time you miss the ball, it resets from a random location. To simulate this, researchers sent random unpredictable signals to the brain tissue whenever it missed the ball.
At first, the brain tissue was moving the paddle in random directions. After all, it was just a small bunch of neurons without the rest of the brain to tell it what to do.
However, over the next few minutes, the brain tissue slowly learned to coordinate its internal electrical activity to avoid missing the ball – sustaining longer and longer rallies with time.
But how? It’s just a slice of brain tissue in a petri dish!
Predictive processing has no issues explaining the results of this experiment. Each layer of neurons is always trying to predict the electrical activity of the neurons in the layer below. Every time the paddle misses the ball, it receives a random unpredictable signal that generates a prediction error. In response, the neurons reorganize their connections to minimize prediction errors in the future. Many rounds of reorganization later, the brain tissue ends up getting connected in a way that minimizes missing the ball.
Now, alternatives to predictive processing can explain the results of the Dishbrain experiment only if we assume that random unpredictable signals somehow count as “punishment” and predictable signals count as “reward”. Yet, they offer no real reason for why this should be the case without invoking predictive processing.
#3 Natural emergence of Self and subjective experience
One of the biggest mysteries of neuroscience is explaining how a bunch of neurons can give rise to the rich subjective experiences of our mind.
Let’s take the colour red. Physics tells us that each colour is related to the wavelength of light that hits our eye. A camera and our eye will both receive the same wavelength.
Yet, we perceive “something more” when we perceive red. After all, a colour-blind person sees the same wavelength but perceives a “redness” that is different from what others perceive. This “something more” is called qualia. It’s the redness of red and the sweetness of sugar. Qualia forms the building blocks of our subjective inner reality.
While many efforts are underway to understand how our brain produces qualia, predictive processing offers the simplest explanation among its rivals.
In predictive processing, the actual wavelength of red light is immaterial to our perception since we never directly perceive the sensations produced by it. Instead, our perceptions come from our brain’s predictions. These predictions are inherently subjective since the brain activity that helps me predict the sensation of red light might differ from the brain activity that helps you predict the same sensation. Thus, qualia.
By the same logic, our brains don’t directly perceive the signals from our own bodies either. Instead, we only perceive the models or simulations our brain creates to predict incoming signals from our bodies. The same goes for our brains. Some regions predict the electrical activity of others. These predictions become the basis for our subjective sense of “self.”
So much explanatory power emerges from a simple assumption: our neurons don’t transmit raw signals, but only the prediction errors that arise from those signals.
Of course, this simplicity alone doesn’t make it right, but it does make it an appealing idea to explore, no?
The Challenges
As you would’ve guessed, we’d already be hailing predictive processing as the new default of neuroscience if it didn’t have challenges waiting for a fix. Let’s look at them in this section, hoping that they may inspire us to attempt solving them.
⚠ Lack of testable predictions
The biggest challenge and opportunity for predictive processing is to define clear, testable predictions that uniquely arise from predictive processing, distinguishing it from other theories. Without such predictions, critics argue that it is still too broad and general, maybe even unfalsifiable in its current form.
⚠ Anatomical implementation
While we now have computational models to implement predictive processing, we still lack a clear explanation for how it may be implemented within the brain. Most of its explanatory power comes from a “black box algorithm” level of understanding.
In its defence, none of its alternatives are any closer.
⚠ Mechanism for Active inference
Active inference, a key component of predictive processing, suggests that organisms don’t just update their priors to reduce prediction errors but also take actions to fulfil predictions (e.g., moving a finger to confirm a prediction that “my finger will move now”). How this can happen remains unanswered.
⚠ Integration with well-established cognitive principles
Predictive processing is yet to be integrated with other well-established cognitive principles, such as reinforcement learning (learning by trial and error with rewards), Hebbian learning (neurons that fire together, wire together) and embodied cognition (cognition is not just something the brain does, it’s something the whole body does).
The Possibilities
Despite its challenges, predictive processing holds the potential to one day become the default theory of how our brains work. Here are the possibilities I’m most excited about –
✦ Unified Framework for Cognition
Starting with just one assumption that our brain operates on the principle of minimizing prediction error, it provides a unified account of how various features of our mind emerge – from perception and action to learning and even higher-level processes like decision-making and consciousness.
If we can develop this into a full-fledged testable theory, it might someday emerge as the “theory of everything” of neuroscience.
✦ A Fundamental Connection with the Physics of Life
The predictive processing framework aligns with the Free Energy Principle, a mathematically grounded theory that explains how biological systems maintain order and resist entropy. This alignment gives it a solid theoretical foundation and connects it to broader principles of thermodynamics and information theory, giving it a strong headstart in explaining why it evolved in the first place.
✦ A New Way to Understand Neurological Conditions
Predictive processing provides new ways to model and simulate several neurological disorders by framing them as disruptions in the brain’s predictive mechanisms. For example, conditions like schizophrenia, autism, and anxiety disorders can be understood as resulting from imbalances in prediction and error correction, offering new avenues for treatment and intervention.
✦ A Gateway to Artificial Sentience
By providing a computational model for how subjective experiences could emerge as a natural side-effect of minimizing prediction errors generated by signals from objective reality, predictive processing opens the door for creating artificial systems with subjective experiences, a sense of self, and possibly even consciousness.
Closing Thoughts
Over the years, I have implemented several forms of predictive processing models: black box algorithms, neural networks, models with altered parameters that successfully reproduce several symptoms of autism and schizophrenia, models that failed at reproducing symptoms of ADHD, anxiety and depression…
It has been an exciting journey, peppered with moments of deep frustration when facing its limitations. But I can confidently say that the deeper I’ve dived into this framework, the more I’ve come to bet on its potential.
So, I hope this post sparks enough excitement in you to explore this beautiful idea on your own. And if you ever need another curious mind to give you company on this journey – hit me up! ✨
