Li Ji-An
1,2,3,
Marcus K. Benna*31 Neurosciences Graduate Program, University of California San Diego, La Jolla, CA, USA
2 Department of Psychology, New York University, New York, NY, USA
3 Department of Neurobiology, University of California San Diego, La Jolla, CA, USA
*Email:
[email protected]IntroductionBackpropagation has been highly successful for training artificial neural networks. However, whether the biological brain implements any variant of backpropagation remains unknown. A key challenge concerns the capability of a single biological neuron to simultaneously encode and transmit feedforward predictions and feedback errors with minimal interference. Here, we propose a neuronal frequency multiplexing framework to address this challenge.MethodsEach model neuron has multiple compartments that multiplex signals in the frequency domain. One dendrite acts as a low-pass filter, and extracts feedforward prediction signals from the low-frequency, direct-current components of the inputs. Another dendrite acts as a high-pass filter, and extracts feedback error signals from high-frequency, oscillatory components of the inputs. The soma integrates both signals, transmitting them to other neurons through its firing rate, which consists of a slowly varying prediction component and an oscillatory error component.ResultsWe demonstrate that this frequency multiplexing algorithm using a simple, local learning rule closely approximates backpropagation in fully connected networks trained on the MNIST dataset and in convolutional networks trained on the CIFAR-10 dataset, achieving comparable performance and similar learning speed as a function of the number of training epochs.DiscussionOur framework implements backpropagation-like training of functionally feedforward neural networks using continuously running, recurrently connected neuronal populations that simultaneously encode and propagate both prediction and error signals with minimal interference. This represents a new solution to the long-standing problem of biologically plausible credit assignment, suggesting a potential computational role for oscillatory signals in coordinating synaptic plasticity.
