Excitatory-inhibitory cortical feedback enables efficient hierarchical credit assignment

The error-backpropagation (backprop) algorithm remains the most common solution to the credit assignment problem in artificial neural networks. In neuroscience, it remains unclear whether the brain could be adopting a similar strategy to modify its synapses. Recent models have attempted to bridge this gap while being consistent with a range of experimental observations. However, these models are either ineffective at propagating error signals through several brain areas or require a multi-phase learning process, neither of which are reminiscent of learning in the brain. Here, we build upon prior work introducing Bursting Cortico-Cortical Networks (BurstCCN), a model which solves these issues by integrating biologically-plausible bursting, dendritic feedback and cell-type specific functional connectivity. BurstCCN uses a burst-dependent synaptic plasticity rule and connection-type-specific short-term synaptic plasticity to enable burst multiplexing. In addition, it relies on apical dendrite-targeting (SST) interneurons to maintain E/I balance and facilitate the encoding of error signals. First, we demonstrate that the BurstCCN can effectively backpropagate errors through multiple layers using a single-phase learning process. Next, we demonstrate that the BurstCCN is capable of learning non-trivial image classification tasks (MNIST and CIFAR-10) as well as a reinforcement learning task (CartPole). Finally, we show how Dalean constraints can be introduced, proposing a role for both inhibitory (SST, PV, NDNF) and disinhibitory (VIP) cell-types. Overall, our work suggests that specific excitatory-inhibitory cortico-cortical connectivity with both short- and long-term synaptic plasticity, jointly underlie single-phase efficient deep learning in the brain.