Neural circuit theory and trained recurrent network modeling of rapid learning

点击翻译按钮获取中文摘要

Humans have remarkable ability to acquire a rich repertoire of concepts stored in semantic memory, which can be deplored in “learning to lean” that facilitates rapid new learning or even one-shot learning. Nonhuman animals are also endowed with “learning to learn”; on the other hand, there is evidence that primates but not rodents possess this mental capability. The underlying brain mechanisms are completely unknown and represent a widely open question at the frontier of Neuroscience today. The present computational project, in conjecture with the experimental projects of this application, has the primary goal of elucidating the neural circuit basis of rapid learning. Progress is this research direction will represent a major step forward in bridging nonhuman primate and human neuroscientific understanding of higher cognition. Our modeling approach integrates large-scale circuit modeling of primate brain based on measured mesoscopic connectivity and training recurrent neural networks to perform cognitive tasks. Together with the proposed experiments in this application, we will develop tools to describe and elucidate neural population dynamics in single trials, which is crucial for neurophysiological analysis of rapid learning (even one-shot learning) without averaging over many repetitive trials in a steady state situation. The main hypothesis is that learning to learn depends on the formation of an abstraction of sensori-motor representations, such as that of task structure or “schema”, which is manifested in a shift of neural representation from the hippocampus to the prefrontal cortex; this conceptual representation enables rapid future learning by efficient changes of connection weights within a low dimensional subspace. This hypothesis will be tested using the state space analysis and dimensionality reduction of the recurrent neural network dynamics. Aim 1 will to be to advance a mesoscopic connectivity-based multi-regional neural network model for rapid learning in categorization, flexible sensori-motor mapping and object-location association. The model will be systematically tested and validated by comparison with behavioral data from category learning and associative learning tasks. Aim 2 will be to uncover neural population dynamics and circuit mechanism of rapid learning in single trials, using state-space analysis and identifying a subspace of neural population dynamics as well as a subspace of connection weights that may correspond to the formation of semantic memory. Aim 3 will be to dissect the differential roles of HPC, PFC, PPC and their dynamical interactions underlying rapid learning, by simulating “area lesion” at different time points of a learning process. A spiking network version of our model will enable us to uncover inter-areal dynamical interactions and their role in rapid learning. Advances in this area would not only be important for the Neuroscience of learning and memory, but also have potentially major implications for the future development of AI, and for shedding insights into the brain mechanism of deficits in semantic memory, which is at the core of fronto-temporal dementia.