If we rely only on our senses we would notice we are seeing a small part of an object. Our eyes capture a small patch of colours but we see a bee on a flower. How does the brain construct this image from the minuscule sensory data?One theory is our brains infer the most likely state of the world (coming up with a "best guess") given our current sensory data and our prior knowledge about the world.

Perception is Inference

"I predict myself therefore I am"

What I believe I am experiencing right now is actually what I am experiencing (called a Posterior Belief). So our reality is weighted average of prior beliefs and sensory data.These prior beliefs are acting like the hierarchical levels separating a neural network. These included, starting from the levels closer to the sensory data:<ul><li>sensations</li><li>gesture, voices, happiness, disgust</li><li>drawings, motions</li><li>memory, math, faces</li><li>reasoning, concepts</li></ul>

Our Bayesian Brain Hallucinates our Reality

Not only we are hallucinating our world. We halucinate ourselves and we halucinate how our self integrates in the constructed world. This is based on:<ul><li>sensory integration</li><li>our sense of agency</li><li>experience of what it means to be "me"</li><li>perspective</li></ul>Ofcourse, in daily life we are not aware of our hallucinations. Meditation or psychedelics are ways to break through our model of ourselves.

Phenomenal Self Model

The Bayesian brain hypothesises that our mind works like an AI predictive model. We don't experience our world directly, but we perceive it. Meaning we construct it using both our senses and our prior beliefs.If this is true then meditation can act as a way to improve our experience of ourself and of the world. By making us aware of our biases and allowing us to change how we create the world. This is the proposition of Shamil Chandaria, one of the AI pioneers, having worked at DeepMind (now part of Google).

Bayesian Brain & Meditation

It is a hypothesis that our brain works through hierarchical predictive processing:<ul><li>the brain is constantly generating predictions about the world and then comparing those predictions to incoming sensory data</li><li>the brain is organised into a hierarchy of processing levels, each of which generates predictions about the level below it. Ex: the highest level of the hierarchy might generate predictions about the overall scene that we are looking at, while lower levels of the hierarchy might generate predictions about specific objects or features in the scene.</li></ul>

The Bayesian Brain Model

How does our brain perceive the world based given the noisy and incomplete sensory data?<ol><li>We have the World creating events.</li><li>A small % of that reality creates sensory data.</li><li>As Agents in the world, we perceive the World. Meaning ur brains have to come out with the best hypothesis for the features of the world given the sensory data.</li></ol>Bayesian theorem is the optimal mathematical solution for this mathematical problem. A very computationally explosive problem. We can get an inspiration from AI systems that are addressing it.

Bayesian Perception Problem

Kant, the famous philosopher, postulated we never have an experience of the "real" world (the noumenal word). Our cognition does not conform to the world, but rather the world conforms to our cognition.We never see the tree as it really is but rather experience our perception of the tree.Kant argued that our minds are not simply passive receptacles for sensory information. Instead, we actively construct our experience of the world by filtering it through our hyper-priors (innate cognitive predispositions). Like our concepts of space & time, our categories & our moral principles.

Kant's Phenomenal World

Neural Networks (which are heavily used in AI) appear to be the solution to the complexity of the Bayesian Perception Problem. They break the problem in levels.You feed a collection of images of faces to a neuronal network and it predicts:<ul><li>low-level features: edges and corners</li><li>mid-level features: noses or eyes</li><li>high-level features: facial details</li></ul>Since this hierarchical predictions are computationally efficient, we can think our mind also solves the complexity of a Bayesian prediction by breaking up the problem in hierarchical levels.

Neuronal Networks solution to the Bayesian Problem

Meditation allows us:<ul><li>to over-emphasise our senses (as opposed to our priors) through attentional meditations (where we count our breaths for example)</li><li>to become aware of our priors (higher level models in the Bayesian brain like our ethical foundations) and the ways we build reality through deconstruction type meditations like mindfulness</li><li>to reconstruct ourselves through constructive meditations like love&kindness.</li></ul>

Meditation as Model Mind Hack

If we place the 3 types of meditation of on a 3D chart we can plot the evolution of meditator from "monkey mind" to enlightenment. Just like the Jhanas, states of deep mental concentration and absorption that are cultivated in Buddhist meditation.In short meditation:<ul><li>lowers the weight of the priors compared to our senses</li><li>attenuates our high-level priors: beliefs, attitudes etc</li><li>allows us to tweak our change our priors or create new ones</li></ul>

Phenomenal Dimension of Meditation

Shamil is a former DeepMind employee. Based on his AI experience he proposes that our minds are operating similarly to an AI model.

A predictive processing account of radical changes in the character of phenomenal experience.  Talk delivered on the 11th October 2022, University of Oxford.
Centre for Eudaimonia and Human Flourishing, Linacre College, University of Oxford, UK
Dr Shamil Chandaria is a senior research fellow at the Centre for Eudaimonia and Human Flourishing.