CCN Seminar with Tri Minh Nguyen (Harvard)
Please join us for a CCN Seminar with Tri Minh Nguyen, postdoctoral associate in the Lee Lab at Harvard Medical School.
Title: Cerebellar connectomics reveals noise resilient learning through structured connectivity
Abstract: The cerebellum is well known to be critical for motor error corrections and classical conditioning where a neutral stimulus is conditioned to predict a positive or negative outcome. Computations supporting these behaviors must be low latency and sensitive to small input differences but also resilient to noise. Theories of cerebellar information processing posit that these computations are performed by two distinct neuron layers: the first layer expands the dimensionality of the input information while the second layer learns a linear decoding (e.g., linear/logistic regression) based on an external teaching signal. These theories, however, have largely assumed random and dense connectivity where, in principle, random network connectivity in the first layer maximizes encoding capacity, and dense connectivity in the second layer maximizes memory capacity. Here, we tested these assumptions using large-scale transmission electron microscopy (EM) and automated segmentation to comprehensively examine network connectivity in the mouse cerebellar cortex. Our results demonstrate that neither layer maximizes capacity: rather, we find evidence of redundant, selective, and sparse connectivity. Using numerical simulations based on this connectomic graph, we show that these non-random connectivity motifs increase robustness of pattern separation and pattern recognition. These results unveil new principles of cerebellar network architecture with implications for artificial neural network design.
Organizer: Center for Computational Neuroscience