Blake
Richards
Department
of
Cell
&
Systems
Biology
University
of
Toronto
Deep Learning with Pyramidal Neurons
Deep learning has led to significant advances in artificial intelligence, in part, by adopting strategies motivated by neurophysiology. However, it is unclear how deep learning could occur in the real brain, due to the difficulty of performing credit assignment without backpropagation. Here, we show that deep learning can be achieved in a biologically feasible simulation by moving away from point neuron models and towards multi-compartment neurons. Like neocortical pyramidal neurons, neurons in our model receive feedforward sensory information and higher-order feedback in electrotonically segregated compartments. Thanks to this segregation, the neurons in different layers of the network can coordinate local synaptic weight updates to achieve global optimization. As a result, the network can take advantage of multilayer architectures---the hallmark of deep learning. This work demonstrates that deep learning can be achieved using segregated dendritic compartments for feedforward and feedback information, which may help to explain the dendritic morphology of neocortical pyramidal neurons.