The overall aim of our research at the AISys lab is to build the foundations that enable us to synthesize truly autonomous systems. To that end, we build novel algorithmic and theoretically principled methods that are grounded in mathematics that enable us to learn concepts and policies via appropriate representations that can enable systems to act rationally and transfer their knowledge and experience across different tasks and environments.
- Transfer Learning
- We develop principled methods to enable learning transferable concepts that promote transfer learning across tasks and environments.
- We develop non-convex, typically multi-objective optimization by incorporating a variation of transfer learning to solve systems problems (e.g., optimize performance) with low sample sizes, because they are expensive, sometimes even impossible.
- We develop principled methods to optimize performance of highly-configurable systems across the stack (software, middleware, hardware).
- Causal AI: Structure Learning and Transfer Learning
- Causal inference provides a set of tools and principles that allows us to reason about questions of counterfactual nature by combining data and structural invariances about the environment — i.e., what would have happened had reality been different, even when no data about this alternate world is available.
- We develop theoretical and practical approaches to solve several difficult challenges in systems including interference and causal debugging and optimization of systems.
- Adversarial Machine Learning
- With the overall goal of building reliable and secure machine learning systems, we develop methods to enhance model robustness and shield such systems against adversarial attacks.
- We are excited about our new initiative on synthesizing dynamic defenses (under patent filing) against adversarial attacks at deployment time.
- Autonomous and Adaptive Systems
- We aim to develop the next generation of autonomous systems (on-device, embedded, heterogeneous, cloud, robotics) that can perceive, reason, and react to complex real-world environments and users with high levels of precision and efficiency.
- To build such fully autonomous systems, we rely on continual knowledge learning and transfer.