Talk by Prof. Tomoki Fukai (CSN Virtual Seminar)
We hereby announce the next talk in 'CSN Virtual Seminar' series
Cognitive computations with self-supervised predictive learning
Speaker: Prof. Tomoki Fukai, Neural Coding and Brain Computing Unit, Okinawa Institute of Science and Technology, Tancha 1919-1, Onna-son, Okinawa 904-0495, Japan.
Abstract
Prediction is a fundamental computation of the brain's neural circuitry. I will show a self-supervising predictive learning rule, which enables single neurons and their networks to perform complex computations such as unsupervised sequence segmentation and blind source separation. Then, I generalize this learning rule to excitatory and inhibitory neurons in a recurrent network model and show that this network model can encode memory traces of multiple probabilistic sensory events into spontaneous activity. Sensory stimuli generate cell assemblies that remember the occurrence probabilities of these stimuli, and the self-organized network replays the learned probabilistic events in spontaneous activity. Within-assembly, but not between-assembly connections, play a crucial role in these replay events, indicating that the underlying mechanism differs from the conventional Markov-chain transitions. Finally, I present a biologically plausible, Ca2+-based model of self-supervised predictive learning rule. In particular, a recurrent network of this model neuron can learn salient input patterns only with a few presentations if, as suggested in experiments, the network contains pre-configured cell assemblies.