The Artificial Intelligence Podcast with Lex Fridman – Matt Botvinick: Neuroscience, Psychology, and AI at DeepMind

Matt Botvinick is the director of neuroscience research at DeepMind.

My sense of wonder comes not from the distant, mysterious stars, but the extremely, intimately close brain.

Information Processing

• Neurons communicate through neurotransmitters that are emitted between each other that change the voltage of the neurons to a point where an electrical signal is eventually released
• There’s still uncertainty about whether that’s an adequate description of how information is transmitted within the brain.
• There are studies that suggest that the precise timing of spikes matters, and there are studies that suggest that there are computations that go on within the dendritic tree that are quite rich in structure, but that really don’t equate to anything that we’re doing in our artificial neural networks.
• Currently, the AI community views the activity rate of their neural networks as analogous to the spike rate of neurons within the nervous system

Meta learning

• Meta learning relates to the old idea in psychology of learning to learn, situations where you have experiences that make you better at learning something new
• It’s like how learning new languages is hard at first, but you eventually anticipate what you need to learn so you can learn more languages more easily
• These learning dynamics also resemble the activation patterns of the prefrontal cortex that learns based on the reinforcement mechanisms of rewards from certain behaviors 
• The question that arises from this is how the prefrontal cortex was programmed to learn from these methods 

A blind spot for at least robotics is human-robot, human-agent interaction.

The prefrontal cortex

• The prefrontal cortex is the cerebral cortex (outer brain layer) region around your forehead that is associated with planning, decision making, and personality expression
• The prefrontal cortex is involved in controlled, willful thoughts and behaviors rather than our habitual, automatic thoughts and behaviors
• Researchers and doctors noted that the brain-damaged veterans of WW1 who weren’t completely incapacitated and were able to live “normal” lives tended to have damaged prefrontal cortices
• The most significant side-effect for these veterans was that they couldn’t process new information

We don’t understand our brain very much

• We understand much of the high-level functions that the brain performs, but much less so of the physical mechanisms behind how those functions operate.
• Cognitive science, neuroscience, and psychology are all different branches of knowledge stemming from the brain with each field only capturing different aspects of the truth.
• Understanding the foundational truths of the mind will require piecing together all the knowledge we have of the brain rather than myopically focusing on neuronal voltage channels or behavioral conditioning.

When you look carefully at functional differentiation in the brain what you usually end up concluding is that the difference between regions is graded rather than being discrete.

Meta-learning, by definition, is a situation in which you have a learning algorithm and the learning algorithm operates in such a way that it gives rise to another learning algorithm.

Understanding the brain contd.

• Cognitive neuroscience today is in a place similar to where genetics was after Mendel established the heritability of traits but before Watson and Crick discovered the double-helix structure of DNA
• We knew DNA was involved in the passing of genetic information, but we didn’t know how. The knowledge of the double helix allowed scientists to uncover the mechanical process of replication that drove the passing of genes. 
• Today, we know the interactions between neurons drives our thoughts and behaviors, but we have yet to understand the exact causal processes behind how electrical signals eventually turn into philosophical musings on consciousness

It’s made most clear in reinforcement learning where you can only learn as much as you can simulate.

Source