brAIn Seminar - Xuan Ma

— 4:30pm

Location:
In Person and Virtual - ET - Baker Hall 341A and Zoom

Speaker:
XUAN MA, Research Assistant Professor, Department of Neuroscience, Northwestern University
https://scholar.google.com/citations?user=pwfGTnoAAAAJ&hl=zh-CN

Decoding motor control from a nonlinear manifold spanning naturalistic behaviors

Recent advances in computational neuroscience suggest that neural activity for motor control is organized within low-dimensional manifolds, capturing the key patterns of neural covariation underlying diverse motor behaviors. This framework enables the projection of complex, high-dimensional neural recordings into more interpretable subspaces. However, our understanding of these neural manifolds remains limited, primarily because most studies have focused on a few highly constrained and stereotyped motor tasks, presumably capturing only isolated fragments of a broader manifold structure. 

In this talk, I will present our work based on neural recordings from monkeys performing tasks in both a conventional laboratory setting and a less constrained, in-cage environment. Although the tasks were nominally similar, involving various forms of grasp, they were performed in quite different contexts. In the cage, the monkey must maintain a quadrupedal stance, generating more complex proprioceptive inputs from all four limbs, than occurs while seated in the primate chair. We found that the latent representations of neural activity in low-dimensional space formed distinct clusters within the neural manifold, which we identified using unsupervised algorithms. By computing a separate linear decoder for each cluster, we achieved significantly more accurate EMG predictions from M1 activity than was possible with a single global linear decoder trained on data from all behaviors. This piecewise linear approach even slightly outperformed a global LSTM decoder. However, the piecewise decoder, like the global linear and LSTM decoders, failed to predict EMG accurately for held-out tasks. 

We also tried to predict data for points that fell between clusters by combining the outputs of multiple decoders based on the geometric relationships between the clusters; these efforts failed as well. We analyzed pairwise Euclidean and geodesic distances between neural activity samples across different tasks and during hour-long continuous recordings. The discrepancy between these distance measures was relatively small for in-lab tasks but significantly larger for in-cage tasks and even more pronounced in continuous recordings that included spontaneous behaviors. We hope that by further studying the topology of the curved manifolds and the geometric relationships between the clusters using more intuitive computational tools, we may be able to build decoders through interpolation across the curved manifold, at least for novel tasks having activity that falls within the convex hull of the training data.    

— 

Xuan Ma is a Research Assistant Professor at Northwestern University. He earned his Ph.D. from Huazhong University of Science and Technology, Wuhan, China, where he worked with Prof. Jiping He to establish a nonhuman primate lab from the ground up. In 2018, he moved to the U.S. and joined Prof. Lee Miller’s lab at Northwestern University as a postdoctoral researcher, investigating how neuronal activity in the brain coordinates hand movements to advance brain-controlled functional electrical stimulation (FES) brain-computer interfaces (BCIs). His research focuses on understanding the neural mechanisms underlying motor control as well as developing advanced computational methods to enhance BCI performance, making them more robust to drifts and adaptable to variations in task demands. 

In Person and Zoom Participation.  See announcement.

Event Website:
https://brain.andrew.cmu.edu/seminar

Add event to Google
Add event to iCal