Tightening the Biological Constraints on Gradient-Based Predictive Coding
This work addresses the biological plausibility of predictive coding models for understanding brain credit assignment and neuromorphic hardware, though it is incremental as it builds on existing gradient-based PC frameworks.
The authors modified a gradient-based predictive coding model to better align with biological constraints, such as positive firing rates and unidirectional synapses, and demonstrated that the modified networks achieve comparable performance to the original model and backpropagation on MNIST data.
Predictive coding (PC) is a general theory of cortical function. The local, gradient-based learning rules found in one kind of PC model have recently been shown to closely approximate backpropagation. This finding suggests that this gradient-based PC model may be useful for understanding how the brain solves the credit assignment problem. The model may also be useful for developing local learning algorithms that are compatible with neuromorphic hardware. In this paper, we modify this PC model so that it better fits biological constraints, including the constraints that neurons can only have positive firing rates and the constraint that synapses only flow in one direction. We also compute the gradient-based weight and activity updates given the modified activity values. We show that, under certain conditions, these modified PC networks perform as well or nearly as well on MNIST data as the unmodified PC model and networks trained with backpropagation.