Recurrent circuits improve neural response prediction and provide insight into cortical circuits

Feedforward convolutional neural networks are currently the state-of-the-art approach for characterizing the transfer function of early visual cortical neurons. It is well known that the visual cortex has massive recurrent connections. Would models using recurrent circuits provide better neural response prediction than feedforward models? Here, we show that recurrent models consistently outperform feedforward ones when their parameters and hyperparameters are carefully matched. These benefits are robust against variation in the type of activation functions, loss function, number of channels, and many other hyperparameters. They are generally higher when lower amounts of data are used, suggesting that recurrence improves data-efficiency. We test our models on three datasets with different stimulus presentation paradigms, and find that the improvement of recurrent models is greater when stimuli are presented through a wide-field aperture as opposed to a receptive field-sized aperture, and for a long duration as opposed to a short one. Since both of those conditions likely increase the amount of recurrent processing involved in generating the recorded cortical data, we view this as supporting the claim that the recurrent models’ learned circuits are performing a similar function to the real cortical circuits. Additionally, we propose a new formulation of our recurrent models as ensembles of multiple, weight-tied feedforward processing paths of different lengths. This reformulation allows us to test the effect of removing paths from the models, and comparing the results between datasets, we find that longer paths, resulting from more recurrent processing, are more important for wide-aperture, long-duration recordings. This adds credence to our analysis based on the recurrent models’ performance improvement, and supports its suggestion that in order to characterize the properties of cortical responses generated by recurrence, stimulation paradigms that evoke that recurrence, as well as models suited to match it, are required.