David Ehrlich: critical periods for wiring the cerebellum скачать в хорошем качестве

David Ehrlich: critical periods for wiring the cerebellum

Deep supervised learning utilizes backpropagation, which attributes errors to connections in a multi-layer network starting at the output layer and working backwards. In contrast, a multi-layer network for biological supervised learning, the cerebellum, receives errors upstream and learns without backprop. Cerebellar learning is simplified by precise connectivity at the output layer, with errors channeled to middle layer neurons with appropriate output layer connections. How is accurate output layer connectivity ensured biologically? We show output layer connectivity is developmentally specified using transparent fish genetically engineered to visualize and manipulate connectivity. Output layer connectivity is sensitive to developmental experience, middle layer and output layer activity, and critically also to the error pathway at the middle layer. Via modeling we show that implementing these developmental influences can afford life-long learning without the need for backprop. However, output layer plasticity must be transient or error attribution is abolished by excessive plasticity and backprop is necessitated. Consistently, output layer connectivity in developing fish was plastic for only a single day empirically, a tight ‘critical period’ when refinement must occur. These data suggest biological networks use error pathways to refine output connections during development, which then enable credit assignment upstream later in life. David Ehrlich studied learning and memory for his PhD at Emory University and a postdoc at University of Toronto before two postdocs in sensorimotor systems at Columbia University and New York University. His lab studies how neural circuits control and learn about coordinated behaviors like swimming and balancing. His lab is in the Integrative Biology Department at the University of Wisconsin-Madison.