Task 009: Multi-modal distributed learning
|Event Date:||January 21, 2021|
|Time:||11:00 am (ET) / 8:00am (PT)
|School or Program:||Electrical and Computer Engineering
Eric Homan, Pennsylvania State University Initial exploration and extension of Double-Pass Networks for distributed and remote sensing systems
The brain has developed feature extractors as well as methods to dynamically retune their receptivity. This behavior can be mimicked in convolutional neural networks to improve performance. The double pass object recognition neural network developed by Itti lab at USC falls into this category with hot swappable Gabor filters based on super classes to facilitate dynamic retuning. At Penn State, the group is exploring extending this network methodology for use in edge and distributed systems. One area of focus is long-term object tracking at the edge, where the problem’s temporal and spatial characteristics provide an interesting use case for these networks. This talk will also present the initial work utilizing RL agents to perform class partitioning during network training. This allows each fine-tuned network to learn weights optimized for its discovered class partition and may provide further opportunities to reduce SWAP at the edge.
Eric Homan is a PhD student working with Vijay Narayanan in the School of Electrical Engineering and Computer Science at Penn State University. His research interests include computer vision, bio-inspired intelligent systems, and hardware acceleration. Eric has performed research on bio-inspired perception systems for the edge, understandable autonomy, and robust autonomous vehicles. He is currently working on object recognition for autonomous systems and information constrained edge perception.