Qiang Liu

Qiang Liu Associate Professor Computer Science University of Texas at Austin lqiang(at)cs.utexas.edu Office: GDC 4.806 ("Qiang" sounds like "Chee-ah-ng", and "Liu" as "l-yo")

Research

I am broadly interested in mathematical and computational techniques for learning, inference and making decision out of data and knowledge.

• Algorithmic foundations: I am interested in developing fundamental yet computationally feasible algorithms for the basic learning, inference, and optimization problems that underpin the cutting-edge AI/ML/statistical technologies. I see myself as a mathematically-minded methodologist who uses advanced math to discover the missing algorithms that are otherwise non-obvious. I prefer simple and elegant methods influenced by Paul Erdos’ “The Book”.

• Applications: I hope the algorithms we develop can help solve real-world problems of societal importance. One aspect of this is to tailor the basic algorithms to have desirable practical properties, such as computational/energy efficiency, robustness and fairness. Another aspect is to develop problem-specific solutions to address the unsolved problems in science and engineering.

<My Google scholar>

UT Statistical Learning & AI Group (to be updated)

Selected Works

• Flow and Diffusion models for Generative Modeling, Domain Transfer, Optimal Transport

  • [ rectified flow (intro, for generative and transfer modeling, for optimal transport, for fast text-to-image generation (new!), ) diffusion bridges and Doob's h-transform (let us build bridges, first hitting diffusion, diffusion w/ informative bridges) ]

• Dynamic Barrier Descent for Optimization/Sampling in the Wild: Constrained, Lexicographic, Multi-objective, Multi-level

  • [ slides, constrained, lexicographic optimization, optimization within Pareto set, finding diverse local modes, bi-level optimization, multi-task learning, sampling on Pareto set, sampling with constrained moments, orthonal-space variational gradient ]

• Steepest Descent Methods for Neural Architecture Optimization: Going Beyond Black Boxes

  • [ splitting steepest descent (intro, paper, slides), firefly neural architecture descent (paper, slides ) ]

• Off-Policy Evaluation

  • [ slides, non-asymptotic confidence intervals, breaking the curse of horizon ]

• Certifiable Robustness and Other Properties

  • [ a functional optimization framework, certified robustness in language models, certified monotonicity ]

• Stein Variational Inference: Approximate Learning and Inference With Stein's Method

  • [ intro, kernel Stein discrepancy (KSD) (intro, paper), Stein variational gradient descent (SVGD) (intro, basic algorithm, SVGD as gradient flow, SVGD as moment matching, gradient free SVGD, matrix kernel), slides, slides, slides, slides, code ]

• Distributed Learning, Information Loss and Curved Exponential Families [paper, slides]

• Reasoning and Decisions in Probabilistic Graphical Models [ thesis, slides, marginal MAP ]

• Variational Inference for Crowdsourcing [paper, slides]

Teaching

• Undergraduate intro to machine learning [here].

• Undergraduate intro to optimization [here]

• Lecture notes on probabilistic learning and inference [here]

• Learning theory (graduate level, scribed notes, not proofreaded!) [here]

• Advanced ML for undergraduates (scribed notes, not proofreaded!) [here]