Effective computations for hippocampal place cell phenomena in sparse untrained random networks

The mammalian brain processes experience-dependent spatial information through poorly-understood network mechanisms thought to depend on particular network connectivity patterns and activity-dependent synaptic plasticity. However, dedicated input connections that learn to shape information about place cannot easily explain many rodent hippocampal place cell phenomena. For example, representational drift notwithstanding, the discharge of each place cell maps to specific locations in a fixed environment, but the discharge of most cells remap to distinct independent locations across environments, despite the fact that most sub-second cofiring relationships amongst hippocampal neuron pairs persist across environments. Whereas some models of hippocampal spatial information processing rely on the dedicated input connections of a for-purpose connectome (ignoring remapping, representational drift, and maintained cofiring), other models use synaptic plasticity implemented by learning rules to alter random input connections, but struggle with either limited capacity, representational drift, and/or biological implausibility. Here, using a randomly tuned network with feedback inhibition, we examine whether the assumptions of a specific connectome and learning-implemented synaptic plasticity are necessary for diverse place cell phenomena. We find that the random network with non-plastic connections accounts for positional tuning, single place fields in small spaces and multiple place fields in large spaces, mixed selectivity, and remapping, amongst other place cell phenomena. This requires excitatory activity to be sparse and organized across stimuli by divisive normalization. Enabling synaptic plasticity only at the network connections (not at network inputs) accounts for additional place cell phenomena including overdispersion, representational drift, and memory tagging. We show by simulations and analytically that DivSparse, a random network with sparsifying inhibition can explain many features of place cell network activity, suggesting that simple biologically-plausible architectures can realize representations of spatial experience that are robust, flexible, and spontaneous.