Task choice influences single-neuron tuning predictions in connectome-constrained modeling

Research efforts in mapping the connectivity of neural circuits has led to a wealth of connectomic data, including the full nervous system of the fruit fly. The dynamical characteristics of the neurons in these circuits, however, are still largely unknown. Recent work has demonstrated that connectome-constrained models with ethological task constraints generate accurate predictions of single-neuron response properties. Specifically, models with connectivity derived from the fruit fly visual system were trained using deep learning to perform a motion vision task. Resulting models accurately predict the known separation into ON and OFF pathways of motion detection and provide testable hypotheses for single-cell function across the whole circuit [1]. However, only one training task and dataset was used—optic flow estimation from Sintel [2]—and the impact of this choice of task on model predictions remains unexplored.

We empirically investigate how task choice influences single-neuron tuning predictions in connectome-constrained modeling by training on a variety of visual tasks and evaluating their predictions. We found that connectome-constrained models are able to learn many different visual tasks, but only tasks involving motion detection reliably yield accurate predictions of direction selectivity curves. Further, we parameterized synthetic optic flow tasks to control the visual scene statistics of the task dataset and found that direction selectivity curves are robustly predicted across a range of dataset statistics.

Together, our results suggest that task constraints relevant for the biological circuit, in this case motion detection, lead to the most accurate single-neuron tuning predictions compared to a variety of alternative task constraints. While we expected that biological realism of dataset statistics would improve predictions, we found that major predictions could be robustly recapitulated across a range of parametrically varied dataset statistics.