A Bayesian-Optimized Convolutional Neural Network to Decode Reach-to-Grasp from Macaque Dorsomedial Visual Stream

Abstract

Neural decoding is crucial to translate the neural activity for Brain-Computer Interfaces (BCIs) and provides information on how external variables (e.g., movement) are represented and encoded in the neural system. Convolutional neural networks (CNNs) are emerging as neural decoders for their high predictive power and are largely applied with electroencephalographic signals; these algorithms, by automatically learning the more relevant class-discriminative features, improve decoding performance over classic decoders based on handcrafted features. However, applications of CNNs for single-neuron decoding are still scarce and require further validation. In this study, a CNN architecture was designed via Bayesian optimization and was applied to decode different grip types from the activity of single neurons of the posterior parietal cortex of macaque (area V6A). The Bayesian-optimized CNN significantly outperformed a naïve Bayes classifier, commonly used for neural decoding, and proved to be robust to a reduction of the number of cells and of training trials. Adopting a sliding window decoding approach with a high time resolution (5 ms), the CNN was able to capture grip-discriminant features early after cuing the animal, i.e., when the animal was only attending the object to grasp, further supporting that grip-related neural signatures are strongly encoded in V6A already during movement preparation. The proposed approach may have practical implications in invasive BCIs to realize accurate and robust decoders, and may be used together with explanation techniques to design a general tool for neural decoding and analysis, boosting our comprehension of neural encoding.