Cellular and subcellular specialization enables biology-constrained deep learning.

Learning and memory in the brain depend on changes in the strengths of synaptic connections between neurons. While the molecular and cellular mechanisms of synaptic plasticity have been extensively studied experimentally, much of our understanding of how plasticity is organized across populations of neurons during task learning comes from training artificial neural networks (ANNs) using computational methods. However, the architectures of modern ANNs and the algorithms used to train them are not compatible with fundamental principles of neuroscience, leaving a gap in understanding how the brain coordinates learning across multiple layers of neural circuitry. Here we leverage recent experimental evidence to test an emergent theory that biological learning depends on specialization of distinct neuronal cell types and compartmentalized signaling within neuronal dendrites. We demonstrate that multilayer ANNs comprised of separate recurrently connected excitatory and inhibitory cell types, and neuronal units with separate soma and dendrite compartments, can be trained to accurately classify images using a fully biology-compatible deep learning algorithm called . By adhering to strict biological constraints, this model is able to provide unique insight into the biological mechanisms of learning and to make experimentally testable predictions regarding the roles of specific neuronal cell types in coordinating learning across different brain regions.