Decision Making for Contextual Neuroprosthetic Control with Dendritic Gated Networks

Laura C. Petrich, Heather E. Williams, Matthew E. Taylor, Patrick M. Pilarski

Abstract: In this work, we examine how an emerging computational neural network architecture—Dendritic Gated Networks (DGNs)—may allow machine agents to detect and accommodate changing contexts while learning. Specifically, we study how DGNs impact real-time, sequential decision making by a machine agent as it supports a human user in their control of a neuroprosthesis. Machine learning algorithms for upper-limb prosthetic control are required to make moment-by-moment decisions on how to map biosignals flowing from the human body to movements of the person’s body-affixed robotic device. When the context changes, as it does during activities in a person’s daily home and working life, such mappings can become inappropriate for a new setting and device performance can be compromised. Dendritic gating enables the acquisition of new learned patterns while retaining previously learned patterns, even during continual or ongoing learning. DGN performance was compared against the current clinical and commercial standard, linear discriminant analysis (LDA), on biosignal datasets across four contextual settings known to impact model performance over time: differing arm postures, electrode displacement, varying days of use, and minimal training data. We found that DGNs are able to accommodate for changing contexts and exhibit less inter-subject variability than LDA. Interestingly, we discovered that DGNs are less confident in their incorrect predictions than LDA. When combined with a rejection threshold, this enables a substantial reduction in the number of incorrect predictions made by the model that would be mapped to robotic movements. DGNs appear to combine the benefits of both linear and deep learning to achieve sample efficient training on complex, non-linear data. This work represents the first investigation of DGNs trained on biosignal data in a real-world setting, and our findings suggest that DGNs hold promise for deployment in human-machine shared decision making scenarios.