Sustainable ML Pipelines (2022–)

In this collaborative project, we develop infrastructure, runtime methods, and tools to enable performance optimization of ML inference pipelines by structure reconfiguration, resource adaptations, model switching, and configuration optimization. This project is in collaboration with USC (US), Queens Marry (UK), Roblox (US), Telefonica Research (Spain), and Seldon Technologies (UK). It is one of the projects in my group related to Design for Sustainability in Computing.

AI for Symbolic Math (2021–)

Mathematics is one of the most precious human heritage. To preserve mathematical knowledge even beyond the human era, in collaborations with the University of Tartu, we develop new methods to solve symbolic math tasks (differentiation, integration, PDE, linear algebra, etc.) with machine learning models such as neural networks. The key ideas behind our approach are to exploit the underlying structures (e.g., tree-like structures) behind symbolic math tasks and transform the symbolic math tasks into tasks such as language translations that deep learning models can do well.

• Papers:
• Pretrained Language Models are Symbolic Mathematics Solvers too!, Arxiv (2021)

NASA RASPBERRY-SI (2020–)

In this collaborative project (USC, CMU, York, UArk, NASA JPL/Ames/GRC), we develop technologies that enable learning-based autonomous planning and adaptation of space landers.

Causal AI: Structure Learning and Transfer Learning (2018–)

In this project, we are addressing the following important questions from theoretical and empirical perspectives: (i) How to learn reliable causal structures from data and how to use the learned causal structures for identifying causal invariances across environments? (ii) How to use the causal structure learning and counterfactual reasoning based on causal invariances for systems problems? In particular, we are extremely interested to apply these theoretical advancements for an explainable and guided performance debugging of highly-configurable systems.

• Papers:
• Learning LWF Chain Graphs: A Markov Blanket Discovery Approach, UAI 2020
• Transfer Learning for Performance Modeling of Configurable Systems: A Causal Analysis, AAAI 2019

ATHENA (2019–)

Despite achieving state-of-the-art performance across many domains, machine learning systems are highly vulnerable to subtle adversarial perturbations. In this project, we propose Athena, an extensible framework for building effective defenses to adversarial attacks against machine learning systems.

• Papers:
• ATHENA: A Framework based on Diverse Weak Defenses for Building Adversarial Defense, Arxiv (2020)

FlexiBO (2018–)

One of the key challenges in designing machine learning systems is to determine the right balance amongst several objectives, which also oftentimes are incommensurable and conflicting. For example, when designing deep neural networks (DNNs), one often has to trade-off between multiple objectives, such as accuracy, energy consumption, and inference time. We developed FlexiBO, a flexible Bayesian optimization method, to address this issue.

• Papers:
• , ()

BRASS MARS (2016–)

How do you build a software system that can function for a century without being touched by a human engineer? This is the herculean task being undertaken by the DARPA Building Resource Adaptive Software System’s (BRASS) program. We, at UofSC with several collborators at CMU, have been developing learning mechanisms to be integrated with quantitative planning to adapt/reconfigure robots to overcome environmental changes at runtime.

• Papers:
• Machine Learning Meets Quantitative Planning: Enabling Self-Adaptation in Autonomous Robots, SEAMS 2019
• Model-based Adaptation for Robotics Software, IEEE Software (2019)
• Transfer Learning for Improving Model Predictions in Highly Configurable Software, SEAMS 2017