Brain States Lecture 2: The Neural Code PDF

Brain States Lecture 2 The Neural Code  Recap Brain Perceived reality Reality  Intuition We start by developing some intuition…  Optimal Codes Crowded beach (13 bytes of information)  Optimal Codes (188,000 bytes of information) Optima...

Brain States Lecture 2 The Neural Code  Recap Brain Perceived reality Reality  Intuition We start by developing some intuition…  Optimal Codes Crowded beach (13 bytes of information)  Optimal Codes (188,000 bytes of information) Optimal Codes Face Optimal Codes  Optimal Codes Why is this drastic compression of information possible? Optimal Codes Face Optimal Codes Optimal Codes It’s a face but… More information is needed: a single word can’t capture it. No Word for “Volcano” in 79 AD The cloud was rising from a mountain – at such a distance we couldn't tell which, but afterwards learned that it was Vesuvius. I can best describe its shape by likening it to a pine tree. It rose into the sky on a very long “trunk” from which spread some “branches.” Pliny the Younger (61-113 AD)  No Words for Modern Art Well, it makes me think of…but… Time (October 1, 2012) Droplet (2016) Mimura Chikuhō (1973- ) National Gallery of Victoria, Melbourne July 2022  A Personal Footnote: Life Imitates Art Ecstasy (1938/1987) André Masson (1896-1987) Serotonergic axons advancing National Gallery of Victoria, Melbourne Hingorani, …, Janusonis July 2022 Front. Neurosci. (2022)  Optimal Codes  Before it takes in any external information, the brain knows what is usual and likely, rare but possible, or entirely impossible.  Technically, this means our environment has a deep statistical structure that the brain can take advantage of:  Time is 1D, but space is 3D.  Every day the sun is above the head and has roughly the same size.  Typically, things fall down and don’t rise up.  Smoke always moves away from its source.  A human face has two eyes above the lips, never below them.  This in turn means that the brain can have a detailed model of the relevant reality, before we open our eyes or start navigating it.  The usual and likely can be assigned the shortest code (e.g. a single word or a single neuronal spike).  The model is adjusted with experience, but we may not have conscious access to it. For example, some zero probabilities you had in your 14-year-old model are no longer zero in your current model.  Brain Code Is Flexible  Professional training can be though of as building an optimal brain code for highly specialized environments (that other people have little experience with).  A rookie is more likely to make errors not because of their lack of knowledge, but because the probability distribution in their model is too unrefined and “flat.”  The rookie’s brain can get lost among many possibilities. A more experienced brain may immediately “see” a single, best solution. Some of these solutions take milliseconds and may not be accessible to the conscious mind. This is why learning from textbooks can go only so far.  Beyond Intuition We now attempt to make these ideas more precise… The Natural Environment has a Quantifiable Statistical Structure A Single Spike Can Carry Rich, Model-based Information READER: Single Spike-Based Coding vs. Rate-Based Coding  Optimal Codes Appear Noisy An optimal code will look like random noise to an uninitiated observer. This may sound counterintuitive, since we often associate information with order. However, long strings of similar symbols or patterns suggest suboptimal coding. Consider the following string: 011101101010101010101010101 Obviously, a more efficient code would be 0111011[10×01] But this code will appear more random (and whimsical) to an observer. Coding in a Biological (Noisy) System Which of the elements are more appropriate for information coding? Coding in a Biological (Noisy) System The system is optimally coded but looks like pure noise!  Coding in a Biological (Noisy) System This is a crucial result, since it means that if one were recording from a neuron that optimized transmission from one point in the brain to another, its spike trains would look like complete junk! This is a cautionary tale, since it tells us that the observation of highly random spike trains might indicate extremely noisy neurons or it might indicate optimal coding. Spikes: Exploring the Neural Code  Optimal Information Transmission in Noisy Neurons?  Radio static and shower noise approximate the theoretically extreme noise known as “white noise.”  To achieve optimal transmission, neural systems may perform “noise whitening.”  This effectively means that meaningful signals may be transmitted through less noisy frequencies.  The result is that now everything looks like pure noise (if you don’t have the key). The Problem the Brain is Solving The Bayesian Model https://en.wikipedia.org/wiki/Bayesian_(yacht)  These probabilities should be built Neural response (R) into the model Prob Objects A B C 0.50 Human 0.5 0.4 0.1 Objects (O) 0.49 Monkey 0.3 0.1 0.6 0.01 Zombie 0.8 0.1 0.1 REALITY BRAIN Classical experimental approaches The problem the brain is actually solving  Neural response (R) Prob Objects A B C 0.50 Human 0.5 0.4 0.1 Objects (O) 0.49 Monkey 0.3 0.1 0.6 0.01 Zombie 0.8 0.1 0.1 Prob(A) = 0.50 × 0.5 + 0.49 × 0.3 + 0.01 × 0.8 = 0.405 Prob(B) = 0.50 × 0.4 + 0.49 × 0.1 + 0.01 × 0.1 = 0.250 Add up to 1 Prob(C) = 0.50 × 0.1 + 0.49 × 0.6 + 0.01 × 0.1 = 0.345  Neural response (R) Prob Objects A B C 0.50 Human 0.5 0.4 0.1 Objects (O) 0.49 Monkey 0.3 0.1 0.6 0.01 Zombie 0.8 0.1 0.1 Prob(Human|A) = Prob(A|Human) × Prob(Human) ÷ Prob(A) Prob(Monkey|A) = Prob(A|Monkey) × Prob(Monkey) ÷ Prob(A) Prob(Zombie|A) = Prob(A|Zombie) × Prob(Zombie) ÷ Prob(A) Prob(Human|A) = 0.5 × 0.50 ÷ 0.405 = 0.62 I see a HUMAN Prob(Monkey|A) = 0.3 × 0.49 ÷ 0.405 = 0.36 Prob(Zombie|A) = 0.8 × 0.01 ÷ 0.405 = 0.02  Neural response (R) Prob Objects A B C 0.05 Human 0.5 0.4 0.1 Objects (O) 0.05 Monkey 0.3 0.1 0.6 0.90 Zombie 0.8 0.1 0.1 What if something changed the brain’s model of what the environment MUST BE, without affecting neuronal responses to real objects? Prob(Human|A) = 0.03 Check calculations Prob(Monkey|A) = 0.02 Prob(Zombie|A) = 0.95 I see a ZOMBIE Prob(O) in Schizophrenia The article is not in the Reader and is not required.  Knowledge is Money  Manipulate the viewers’ Prob(O).  Let them convince themselves they can see or hear it too. 2004 – 2016 The Brain’s Inferences are Perceptually Real What is the assumption the brain is making? The Brain’s Inferences are Perceptually Real READER: Laeng et al. (2022) The Brain’s Inferences are Perceptually Real In normal observers, gazing at one’s own face in the mirror for a few minutes, at a low illumination level, triggers the perception of strange faces, a new visual illusion that has been named ‘strange-face in the mirror’. Individuals see huge distortions of their own faces, but they often see monstrous beings, archetypal faces, faces of relatives and deceased, and animals. […] READER: Caputo (2013) The Brain’s Inferences are Perceptually Real https://www.youtube.com/watch?v=rb1CZfUumDI The Brain’s Inferences are Perceptually Real Illusions with an assumption switch (see the posted slide show)  Markov Blankets The Bayesian model can be extended to a more sophisticated framework that introduces the so-called Markov blankets. In this framework, a number of elements (”particles”) separate causally from other elements and begin to appear as if they were the “inside” of one entity. The rest of the elements become the “outside.” The “inside” subset is a Markov blanket. Markov blankets can be inside other Markov blankets. This may explain cells, organs, and higher-order systems.  Markov Blankets Dr. Karl Friston has proposed that the fundamental description of the brain should be based on Markov blankets (Friston et al., 2021).  Markov Blankets Dr. Friston is one of the most highly cited living scientists and in 2016 was ranked No. 1 by Semantic Scholar in the list of top 10 most influential neuroscientists. (Wikipedia) He also writes faster than Stephen King and can be uniquely incomprehensible. (Dr. Janusonis) The concept of the Markov blankets was introduced by Dr. Judea Pearl, a highly innovative computer scientist. He gave a talk at UCSB in 2014. He notes that the setup of most statistical tests is fundamentally incorrect (e.g., they assume that the system’s covariance structure is not affected by an experiment). A side note: Daniel Pearl, the journalist brutally murdered in 2002, is Dr. Pearl’s son. Markov Blankets: The Human Story Karl Friston Judea Pearl Daniel Pearl (1959- ) (1936- ) (1963-2002 ) University College London UCLA