Visual processing in the retina has been studied in great detail at all levels such that a comprehensive picture of the retina's cell types and the many neural circuits they form is emerging. However, the currently best performing models of retinal function are black-box CNN models which are agnostic to such biological knowledge.

Additional Feedback: Is there any way to find the same phenomenon shown for the BCN in the LSTM? The BCN is a nice step towards merging biophysical models with backprop and gradient descent training routines, but I think it is still taking advantage of a large amount of retina-specific knowledge. Perhaps there's a way to identify the ersatz representation of bipolar and amacrine cells within the LSTM? I understand there's no cell types engineered into this model, but perhaps guided by knowledge of its specific computations (i.e., gates could act as a stand in for ACs), or a clustering analysis of its unit responses, you could identify a correspondence between its parameters and those in the neural data? Demonstrating that the LSTM representations are too idiosyncratic or entangled to do such a thing would strengthen the argument in the discussion that "...such predictions would not be easily possible from a pure systems identification approach."

Visual processing in the retina has been studied in great detail at all levels such that a comprehensive picture of the retina's cell types and the many neural circuits they form is emerging. However, the currently best performing models of retinal function are black-box CNN models which are agnostic to such biological knowledge. Here, we present a computational model of temporal processing in the inner retina, including inhibitory feedback circuits and realistic synaptic release mechanisms. In pharmacology experiments, the model replicated in silico the effect of blocking specific amacrine cell populations with high fidelity, indicating that it had learned key circuit functions. Also, more in depth comparisons showed that connectivity patterns learned by the model were well matched to connectivity patterns extracted from connectomics data. Thus, our model provides a biologically interpretable data-driven account of temporal processing in the inner retina, filling the gap between purely black-box and detailed biophysical modeling.