[PAST EVENT] Colloquium talk: Ashley Gao
LocationMcGlothlin-Street Hall, Online on Zoom
251 Jamestown Rd
Williamsburg, VA 23185Map this location
Deep learning (DL) models are usually deployed in cyber physical systems (CPSs). However, DL models are traditionally trained on samples that are clean and free of real-world realisms, such as distortions. To bridge the gap between the clean, undistorted training samples and the dirty, distorted testing samples, three categories of solutions are proposed. First, we can directly add in the realisms to the training samples if we know what realisms to be encountered. We demonstrate the efficacy of this category of solutions through an emotion detection classifier and a verbal conflict detection classifier that we deployed in six real households. Each deployment lasted for four months, and we evaluate the performance of the classifiers using the collected data that have real-world realisms in it, a.k.a. reverberation effects, background noise, and deamplification effects. The second type of solutions is more algorithmic, which propose to domain-adapt using unsupervised domain adaptation (UDA) from the clean source domain of the training data to the dirty target domain of the testing data. We demonstrate the performance of one of our novel UDA algorithms, the MiddleGAN, using standard UDA benchmarks. The third category of solutions incorporates world knowledge into the training process of DL models so the models can be robust towards the realisms. Graph neural networks (GNNs) are a way to incorporate world knowledge if the data can be represented as graphs. We propose a novel GNN-based algorithm, titled D-A3T-GCN, which incorporates the diffusion theory and attention-based GNN to predict traffic volume and speed using standard benchmarks on traffic prediction.
Ashley Gao is a fourth-year PhD student in the Department of Computer Science at the University of Virginia (UVa), advised by Professor John A. Stankovic. Before joining professor Stankovic’s group, she received her M.S. in Computer Science from the UVa. She also received the B.S. degree in computer science and the B.A. degree in literatures of the world from the University of California, San Diego (UCSD). Her research direction lies in the field of transfer learning and domain adaptation, as well as the intersection of deep learning and smart health.
Zoom link: https://www.cs.wm.edu/zoom