DS Ph.D. Qualifier Presentation | Quincy Hershey | Thursday, January 18th @ Noon | Zoom

DATA SCIENCE 

Ph.D. Qualifier Presentation

Quincy Hershey

 Thursday, January 18th, 2024 | Noon - 1:00PM EST

Zoom: https://wpi.zoom.us/my/rcpaffenroth

Committee: 

Randy Paffenroth, PhD Advisor, Mathematical Sciences, WPI

Roee Shraga , Co-Advisor, Computer Science, WPI

Yanhua Li,  Co-Advisor, Computer Science, WPI

Title: Reinforcement Learning with Sparse RNNs

Abstract: 

This research uncovers primary drivers in optimizing Recurrent Neural Network (RNN) network architectures based upon effects of the input and output dimensions on the dimensions of the resulting weight matrices.  These findings generalize beyond recurrent neural networks to the broader realm of neural networks.  As a result, practitioners may better predict expected model performance and optimize model architectures PRIOR to having conducted training.  Sparsity within the weight matrix plays a key role in both developing understanding of model characteristics and allowing optimization of model architectures.  Sparse RNNs are benchmarked against LSTM models on a series of tasks including offline reinforcement learning and anomaly detection with optimized RNNs demonstrating a clear advantage over LSTMs owing to more inherent flexibility.