Trustworthy Deep Learning

Undergraduate

In recent years, deep learning has seen applications in many fields, from science and technology, to finance, humanity, and businesses. However, real-world, high-impact machine learning applications demand more than just model performance. In particular, deep learning models are often required to be “trustworthy,” so that domain experts can trust that the models consistently behave in a way that corresponds to their domain knowledge. For example, medical experts would expect a deep learning diagnosis model to be able to explicitly utilize medical domain knowledge in its prediction; an insurance company would expect a decision on insurance price to be explainable in terms of risk factors; a financial company would expect its fraud detection model to be robust to adversarial attacks; a physicist would expect models to provide consistency with the underlying laws. This course introduces various fields of trustworthy deep learning, including model robustness, defenses for adversarial attacks, interpretability, explainability, fairness, privacy, domain adaptation, rules, and constraints. The course covers some of these aspects in the context of graph neural networks but also covers many other ML models in general deep learning, natural language processing, and computer vision. Prerequisites: a course in linear algebra and multi-variable calculus. Familiarity with PyTorch and other common Python libraries such as Numpy, Sklearn. Deep learning courses such as CPSC 452 or 453 are recommended. TTh 4pm-5:15pm