Continuous State Estimation With Synapse-constrained Connectivity

Continuous state estimation is a fundamental problem addressed by neural systems which underlies complex capabilities such as navigation. While studies in heading direction estimation of the fruit fly D. melanogaster have uncovered how this computation can be performed in a ring attractor architecture, it is unclear how additional synapse-level architectural details contribute to the functional performance of heading direction estimation. In this work, we find a consistent, repeated, motif in synapse-level connectivity data which captures the connectivity distributions of connections between all investigated neurons in the ring attractor. This identified motif contains novel, additional synapses to adjacent motifs that differ from previous theoretical models and anatomical studies. To investigate the properties of this identified highly-recurrent architecture in computational studies of angular path integration, we develop a parameterized backpropagation-through-time approach to train the attractor based on both observed experimental activation patterns and heading direction estimation accuracy. We find that a synapse-constrained network model with the identified additional directional offsets can perform angular path integration, but does not have improvements in accuracy in relation to models with null comparison architectures, which suggests potential additional functional roles of this connectivity structure. Our work presents a proof of principle for how detailed synapse-level motifs can be investigated via network optimization based on both experimental observations and functional performance as well as furthers a detailed understanding of continuous state estimation in biological systems.