Triple equivalence for the emergence of biological intelligence

Hello, welcome. It is July 3rd or 4th, depending on where you are, and we are in Active Inference Guest Stream 110.1 with Takuya Isomura discussing the triple equivalence for the emergence of biological intelligence. Previously, we had some discussions in the Livestream 51 series when we visited Area 51 together, and we talked a lot about a double equivalence, and now the bar has been raised, the increment has been incremented, and this is a really exciting area that I'm sure we'll have a lot to discuss and learn from. So thank you again for joining, and looking forward to the presentation. Great. Thank you, Daniel, for the invitation. Thank you for inviting me for this guest talk of Active Inference Institute. I'm very happy to share my recent paper on the triple equivalence. So let's get started. So I'm interested in the theory of intelligence in general. So maybe the baby has no such high intelligence, but through learning, it obtained the intelligence. This can be modeled or understand in terms of the self-organization of neural network. So newborn neural network may don't have may not have enough information regarding the external world, but through some optimization, which is typically expressed in the world. So the energy minimization. It achieves some It achieves some It achieves some Obtaining Obtaining of some Good structure to Recapitulate the external world, which is the basis for prediction and making insight or making creativity. So we would like to model those emergence of intelligent function within the neural network. Through the NMV energy minimization. So far, there are. Energy minimization. So far there are. To my opinion, there are some Three major pillars in the theory of intelligence. The theory of intelligence. So far, there are three major pillars in the theory of intelligence. So intelligence. So intelligence. So intelligence. So intelligence. So intelligence. Occurs in the brain. So we Model the brain using the network of neural ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail ail And this is a diagram of electrical circuit. And similarly, a single neuron can be modeled like this. So this diagram represents a dynamical equation of a single neuron. And through some approximation, we can represent such a dynamics in terms of the energy minimization. So here, the dynamics that descent, this energy gradient, energy landscape represents the time change of the neural activity like this. So this is a view from a dynamical system. Another view of intelligence is the computation. So it is well known that any algorithm can be expressed in terms of the Turing machine. Turing machine is a simple abstract machine that expresses the computation, which involves a finite state machine that makes some computation. And there is a long tape. So in the tape, some information is encoded. So like a binary signal. And when this finished state machine reads this information, it incorporates that information into this state machine header. So those components can be expressed by mapping shown here. So given the current state and readout information, it provides the move and writing information and the next state. So this is a basic idea…