A01Creation of novel paradigms to integrate neural network learning with dendrites
Taking recent findings in brain science into account, we will explore a novel architecture and learning theory for creating brain-inspired artificial intelligence. To this end, we aim to integrate dendritic information routing of bottom-up and top-down signals in neural network-level computation. Specifically, we will mathematically and biologically extend the self-consistent surprise minimization, our recent theoretical proposal for dendritic learning. This learning rule hypothesizes that neurons minimize a mismatch between dendritic synaptic input and somatic spike output to detect temporal features. We apply our dendritic neuron models to difficult biological and engineering problems including blind source separation of correlated signals, analysis of large-scale neural activity data, and construction of cortical circuit models for hierarchical Bayesian computation. Our goal is to find out novel principles of learning and neural computation useful for multi-purpose artificial intelligence systems.