I will introduce the cross-disciplinary field of theoretical/computational neuroscience, and its recent opportunities in China. To illustrate the field by examples, I will discuss strongly recurrent neural network models for elemental cognitive functions such as decision-making (how the brain makes a risky choice among several options based on expected outcomes). Moreover, I will argue that we are entering a new era of computational neuroscience, in close interplay with experimental advances, for understanding multi-regional large-scale brain circuits bridging neuroscience with artificial intelligence and psychiatry.
The hierarchical organization of the brain’s ventral visual pathway has inspired the feedforward connectionist architectures used in state-of-the-art deep learning methods that have begun to transform applications as diverse as image recognition, disease diagnosis and self-driving cars. However, it is well-known that there are way more feedback projections than feedforward ones in the brain. The function of these feedback projections remains an open question in neuroscience. In this talk, I will talk about our recent result about categorical perception, a hypothesis regarding how feedback projections can provide abstract category knowledge that is able to alter our sensory perception. According to this hypothesis, category learning would warp our perception such that differences between objects that belong to different categories are exaggerated (expansion) while differences within the same category are deemphasized (compression). This suggests a top-down influence from category-selective to feature-selective representations, but the underlying neural mechanisms have not been established. To gain insight into this question, we examined data from behavioral categorization experiments in non-human primates. In the experiments, monkeys performed the same visual motion discrimination task before and after visual motion categorization training. Data analysis shows that, after categorization training, stimuli within the same category were more difficult to discriminate than before categorization training, while the change for stimuli that belong to different categories was less pronounced, supporting compression without clear expansion. To explain this result, we built a neural circuit model that incorporates key existing experimental findings and makes new predictions, including: (1) learned categories are encoded in the spiking activities of neurons in the lateral intraparietal (LIP) area, (2) neurons in the middle temporal area show graded encoding of stimulus motion directions and (3) neurons in the medial superior temporal (MST) area integrate top-down category and bottom-up motion direction information. This model proposes that it is mainly through the feedback projections from LIP to MST that learned categories induce categorical perception. We find that this prediction is largely consistent with recent single neuron recordings in the MST and LIP areas. Collectively, we show the first behavioral evidence for compression in a visual motion discrimination task in non-human primates and develop a biological neural circuit model that allows us to make experimentally testable predictions, thereby elucidating the possible underlying neural mechanisms of categorical perception.
November. 19, 2018
For more details, please refer to 计算神经科学：大尺度脑网络与反馈投射