Willie Neiswanger
Machine learning at Stanford ⊃ Computer Science ⊃ SAIL / StatsML
I am a postdoc in computer science at Stanford University, working with Stefano Ermon and affiliated with the StatsML Group, Stanford AI Lab, and SLAC.
Research: I develop machine learning methods to perform efficient optimization and experimental design in costly real-world settings, where resources are limited. My work spans topics in active learning, uncertainty quantification, Bayesian decision making, and reinforcement learning. I apply these methods downstream to solve problems in science and engineering, for example in the physical sciences and machine learning systems.
I have also worked on distributed algorithms for scalable machine learning, and I develop/maintain software libraries for multilevel optimization, uncertainty quantification, AutoML, and Bayesian optimization.
Education: I completed my PhD in Machine Learning at Carnegie Mellon University, where I was advised by Eric Xing and collaborated with Jeff Schneider and Barnabas Poczos.
Previously, I studied at Columbia University, where I worked with Chris Wiggins and Frank Wood.
News
- July 28, 2023 Co-organized the Differentiable Almost Everything Workshop at ICML 2023.
- Mar 23, 2023 Co-organized the Modern Adaptive Experimental Design and Active Learning Online Reading Group.
- Jan 20, 2023 New paper on automatic differentiation for multilevel optimization in ICLR 2023 (notable-top-5%).
- Jan 20, 2023 New paper on policy identification for active reinforcement learning in ICLR 2023 (notable-top-5%).
- Jan 20, 2023 New paper on a framework to combine weak supervision and generative modeling in ICLR 2023.
- Jan 1, 2023 New paper on offline imitation learning with suboptimal demonstrations in AAAI 2023.
- Dec 7, 2022 New paper on uncertainty quantification with pre-trained language models in EMNLP 2022.
- Dec 2, 2022 Invited talk at the Workshop on Gaussian Processes and Decision-making Systems at NeurIPS 2022.
- Oct 10, 2022 New paper (+ code) on trajectory information planning for exploration in RL in NeurIPS 2022.
- Oct 10, 2022 New paper on decision-theoretic entropies for generalizing Bayesian optimization in NeurIPS 2022.
- July 22, 2022 Co-organized the Real World Experiment Design and Active Learning Workshop at ICML 2022.
- May 15, 2022 New paper (+ website) on likelihood-free Bayesian optimization in ICML 2022 (long talk).
- May 15, 2022 New paper on a modular conformal calibration framework for UQ in ICML 2022.
- Jan 28, 2022 New paper (+ blog post) on experimental design and reinforcement learning in ICLR 2022.
- Jan 1, 2022 New paper (+ website) on large-scale object counting in satellite images, in AAAI 2022 (oral).
- Oct 15, 2021 New paper (+ code) on quantile methods for calibrated uncertainty quantification in NeurIPS 2021.
- Oct 15, 2021 Two papers on explainable machine learning and personalized benchmarking in NeurIPS 2021.
- July 14, 2021 Our paper on Pollux was awarded the Jay Lepreau Best Paper Award at OSDI'21.
- June 10, 2021 New paper (+ website) on Bayesian Algorithm Execution (BAX) and InfoBAX, in ICML 2021.
- June 1, 2021 I co-organized the Machine Learning for Data (Creation, Privacy, Bias) Workshop at ICML 2021.
- Apr 1, 2021 New paper (+ AdaptDL) on Pollux, a deep learning cluster scheduler/tuner, in OSDI 2021.
- Mar 16, 2021 New paper (+ code) on uncertainty quantification with martingales for GPs in ALT 2021.
- Mar 9, 2021 New paper on active classification for catalyst discovery in the Journal of Chemical Physics.
- Jan 12, 2021 New paper (+ code) on a framework for interactive weak supervision in ICLR 2021.
- Dec 22, 2020 Released Uncertainty Toolbox, for predictive UQ, calibration, metrics, and visualization.
- Dec 2, 2020 New paper (+ code) on BANANAS, a method for neural architecture search, in AAAI 2021.
Projects
- Betty An automatic differentiation library for multilevel optimization and meta-learning.
- Bayesian Algorithm Execution (BAX) Extending Bayesian optimization to computable function properties defined by algorithms.
- Uncertainty Toolbox A toolbox for predictive uncertainty quantification, calibration, metrics, and visualization.
- Naszilla A python library for neural architecture search.
- AdaptDL A resource-adaptive cluster scheduler for deep learning training.
- CASL Project An open toolkit for composable, automatic, and scalable learning.
- ProBO A framework for using probabilistic programming in Bayesian optimization.
- Bayesian Optimization and DOE NASBOT for neural architecture search, MPS for design of experiments, and Dragonfly.
-
Prior Swapping
Efficient algorithms for incorporating prior information, post-inference.
-
Embarrassingly Parallel VI
Communication-free distributed variational inference in nonconjugate models.
-
Embarrassingly Parallel MCMC
Asymptotically exact, communication-free distributed
posterior sampling.
-
Fast Function-based Regression
Fast distribution-to-real and function-to-function nonparametric regression.
-
GPU for Time-varying PYPs
Generalized Polya urn for time-varying Pitman-Yor processes.
-
Parallel Frank-Wolfe Optimization
Asynchronous parallel block-coordinate Frank-Wolfe optimization algorithm.
-
LRO Models for Link Prediction
Latent random offset model for interpretable citation prediction and exploration.
-
DDP Object Tracking and Modeling
Dependent Dirichlet process mixtures for unsupervised object detection and tracking.
-
Cell Motility Analysis
TIAM: the tool for integrative analysis of cell motility.