Graduate Student Seminar Series – Yinghe Sun

Graduate Student Seminar Series
Please ensure you invite your Principal Investigator by adding their email via the ‘Add Guest’ button and they will also be notified of your presentation.
Location: MS2158 – 1 King’s College Circle
Presentation Title: Identifying generalizable features in peripheral nerve recordings for improved neuroprosthetic control
Abstract:
Peripheral nerve interfaces can be used to create advanced assistive technologies. Neural networks associated with multicontact nerve cuff electrodes can selectively record and discriminate neural recordings and facilitate neuroprosthetic control. Due to variations in device positioning and anatomy, neural networks trained on one subject currently cannot generalize to others. To take advantage of available data from other subjects, the objective was to train a neural network whose encoder portion can extract representations that generalize effectively when using transfer learning to adapt the classification to new subjects.
The study applied neural networks to classify naturally evoked compound action potentials corresponding to three different sensory stimuli. The datasets were obtained from the sciatic nerves of 9 Long-Evans Rats through 7×8-channel cuff electrodes. To leverage data from multiple subjects, we pre-trained the network on either one subject or merged data from multiple subjects, then used cross-validation to retrain and evaluate it on a separate target subject. Layer freezing was applied to identify which part of the encoder would best generalize.
Pre-training with merged datasets led to a significant increase in mean macro-F1 score compared to subject-specific models trained from scratch (0.810±0.130 vs 0.733±0.121, p < 0.05), regardless of the number of frozen layers. Pre-training on a single subject did not lead to a significant improvement.
A pre-training approach combining data from multiple subjects shows significant improvement in classification performance. The study developed an encoder that benefits classification performance on unseen subjects despite anatomical variability and device positioning differences.
Supervisor Name: José Zariffa
Year of Study: 3
Program of Study: PhD
Powered by Calendly.com