RESTRICTED BOLTZMANN MACHINES AND THEIR FUNCTION IN CORTICAL RECORDINGS

In neuroscience, the continuous generation of new and more complex datasets raises challenges for modeling the statistical structure underlying neural activity. While several machine learning approaches achieve strong predictive performance, they often rely on complex architectures that make it difficult to interpret the probabilistic structure of the data. Therefore, we decided to develop new models for neural data based on Restricted Boltzmann Machines (RBMs), one of the simplest representation-based generative models. We applied these models to cortical recordings in the mammalian brain. This class of machine offers an interpretative framework providing meaningful features of the neural data. Using calcium imaging recordings from the primary visual cortex (V1) of mice and electrophysiological recordings from the primary and secondary visual cortices (V1 and V2) of non-human primates, we demonstrate that our RBMs (so called Conditional RBMs, Potts RBMs, and Adversarially constrained RBMs) allow explicit incorporation and interpretation of behavioral parameters of the experiments, such as the stimuli, the decision of the animal, the movement of the body… We demonstrate that these generative RBM variants successfully model high-order correlations present in neural activity and enable controlled conditioning and sampling under specified stimuli.