He is also known for his work in theoretical neuroscience, looking at AI and the brain, with theories around the free energy principle and active inference. In 2022 he joined the US cognitive computing company VERSES.
Free Energy Principle
In The free-energy principle: a unified brain theory? (2010) he defined the brain as an organ that tries to minimise ‘free energy’. This is an information theory ‘measure’ of the surprise or prediction errors encountered. By minimising free energy, the brain tries to reduce the difference between its predicted model of the world and incoming sensory inputs. In simpler terms, our brains constantly try to align their understanding or model of the world with the actual information received through our senses. When there is a mismatch between what the brain expects and what it perceives, it adjusts its expectations to reduce this gap. The theory is formulated in mathematics.
Imagine the brain being a captain of a ship sailing through fog. The captain (the brain) tries to minimise surprises or unexpected changes (free energy) by predicting the route ahead as accurately as possible. When the actual sea conditions differ from the captain's predictions, the captain adjusts the course to reduce these surprises. Similarly, the brain constantly updates its understanding to minimise the difference between its predictions and what it actually experiences.
Active Inference
Building on the ‘free energy’ idea, active inference holds that the brain actively samples and acts on the world to gather sensory data that confirms its predictive models, rather than just passively perceiving. Active inference is our constant attempts to try to reduce uncertainty about the world and readjusting between what our brains expect to happen and what actually happens. You may want to return serve in tennis, based on your prediction of where the ball will be relative to your body and racket. As you get better you readjust so that your predictions match the better outcome. This constant loop of predicting, checking, and acting is what Friston's active inference theory is all about.
Inference in conscious perception is an idea that goes back to 1867 with Hermann von Helmholz and his Treatise on Physiological Objects. Our brains are networks of inferences where we solve perceptual, geometric and visual cues, such as we find in faces and reading, similarly in listening and understanding speech. The brain is a connected web of super-fast inferencing, settling on solutions or best-fits, constantly optimising in the light of new data and providing continuity in consciousness.
Bayesian Brain Hypothesis
Friston views perception and cognition as fundamentally Bayesian processes. A Bayesian model for the brain sees the brain process information as a Bayesian inference machine that continuously updates its beliefs about the world based on incoming sensory information and prior knowledge or expectations. The brain has pre-existing beliefs or knowledge about how the world works. These beliefs are formed from past experiences and learning. In Bayesian terms, these are called ‘priors’ and as we interact with the world, our senses provide us with new information. This is the data that the brain uses to update its beliefs. The brain evaluates how likely the new sensory information is, given the current beliefs or hypotheses about the world, then combines the new information with its prior beliefs to form updated beliefs. This process is mathematically described as calculating the ‘posterior’ probability, which integrates the likelihood of the new information with the prior probability. The brain makes further predictions about future sensory input based on these updated beliefs. If the new sensory input does not align with predictions, the brain experiences a prediction error, which it then uses to further update and refine its beliefs.
This Bayesian approach to understanding the brain is part of a broader movement in cognitive science and neuroscience that views the brain as an active inference machine, rather than a passive information processor. It has implications for understanding perception, decision-making, and even psychopathologies. For example, hallucinations in schizophrenia can be thought of as a case where the brain gives too much weight to its priors, leading to distorted perceptions of reality.
He includes another level of theorising around Hierarchical Predictive Coding, namely coding across cortical layers of the brain, with higher levels generating top-down predictions that are matched against bottom-up sensory evidence. In computing, there are processes that are analogous to Hierarchical Predictive Coding. For example, machine learning models, particularly in deep learning, often use a hierarchical structure where layers of processing are used to build up increasingly complex representations of data. Neural networks, for instance, can be trained using backpropagation, which is an algorithm that minimises the prediction error across the layers of the network.
Unified Brain Theory
Friston aims for an overarching theory where processes like perception, action, reasoning and learning all follow from the same core principles of prediction error minimisation. In essence, Friston sees the brain as a hierarchical prediction machine constantly trying to match its predictive models of the world with sensory inputs through Bayesian belief, updating across multiple neural levels. This perspective aims to provide a unified account of different cognitive functions based on a computational model of the brain.
Critique
Karl Friston's theories, including the Free Energy Principle and Active Inference, although influential, are complex mathematical formulations, even for experts to understand. The hypothesis is also very broad, attempting to explain all brain processes and has been criticised for failing to explain the sheer variety of neural and cognitive processes, Some also claim the theory does not actually map onto the actual physiology of the brain. The main criticism is that it is reductionist, reducing all brain functions to mathematical models that ignore the richness of internal subjective experience, language and culture.
Legacy
Despite these critiques, Friston's work remains a unique cornerstone of theoretical neuroscience, stimulating much research and discussion in the field, especially in the light of the success of AI. The debate around his theories reflects the broader challenges of understanding the intricate workings of the brain, enlivened by advances in smart technology.
Discussion
The convergence of cognitive science, neuroscience, brain scanning, mathematics and informed by successes in AI has stimulated ideas around the connected network of the brain an incredibly powerful computer. It is not that the brain does exactly what multimodal models do in terms of being trained, fine-tuned, even altered by other humans correcting our models. What we are seeing are more like strong clues as to how a computational brain could work that matches the experimental evidence in perception and action. Theory and experimental evidence is in its infancy but it may just allow us to escape from the Cartesian idea of the self and mind, defined in terms of fixed cognitive concepts, modules and fictitious unconscious entities.
No comments:
Post a Comment