Predicting proprioceptive cortical anatomy and neural coding with topographic autoencoders

Proprioception is fundamental for the control of movement, its loss producing profound motor deficits. Yet basic questions of how pose and movement are represented, as well as how these representations are arranged across the somatosensory cortex, are unclear. To this end, we adopt a task-driven modelling approach, using a spiking variational autoencoder to approximate a population of cortical neurons. We optimize the model to encode natural movement stimuli derived from recordings of human kinematics and impose biological constraints which we hypothesise to be important for reproducing characteristics of proprioceptive neural coding, namely, enforcing a spike-based code and implementing lateral effects between neighbouring neurons in the model to drive topographical structure in neural tuning. To evaluate the effectiveness of these principles at reproducing empirical observations in neural data (without directly fitting to neural data), we task our model with encoding movement kinematics during a centre-out reaching task and compare activity in modelled neurons to recorded neurons in area 2 of monkeys performing the same task. The model reproduces several key features of neural tuning at both the level of individual neurons, and the spatial organisation of their tuning across the cortical surface. Furthermore, we demonstrate the importance in training on data from the true distribution of natural behaviour, with the model failing to reproduce key properties of the empirical data when trained on stereotyped reaching behaviour only. We then highlight two testable predictions made by the model: 1. The arrangement of directional tuning across the cortex has a blob-and-pinwheel-type geometry. 2. Few neurons encode just a single joint. In summary, the topographic VAE (Author et al, XXXX) provides a principled basis for understanding sensorimotor representations and their spatial organisation in cortex. These basic scientific principles may have application to the restoration of sensory feedback in brain-computer interfaces (Weber et al, 2012).