ML for design

Compositional Generative Inverse Design

Tailin Wu*, Takashi Maruyama*, Long Wei*, Tao Zhang*, Yilun Du*, Gianluca Iaccarino, Jure Leskovec

ICLR 2024 Spotlight
We introduce a method that uses compositional generative models to design boundaries and initial states significantly more complex than the ones seen in training for physical simulations

ML for simulation

BENO: Boundary-embedded Neural Operators for Elliptic PDEs

Haixin Wang*, Jiaxin Li*, Anubhav Dwivedi, Kentaro Hara, Tailin Wu

ICLR 2024
We introduce a boundary-embedded neural operator that incorporates complex boundary shape and inhomogeneous boundary values into the solving of Elliptic PDEs

ML for simulation

Uncertainty Quantification for Forward and Inverse Problems of PDEs via Latent Global Evolution

Tailin Wu*, Willie Neiswanger*, Hongtao Zheng*, Stefano Ermon, Jure Leskovec

AAAI 2024 Oral (top 10% of accepted papers)
We introduced uncertainty quantification method for forward simulation and inverse problems of PDEs using latent uncertainty propagation.

ML for simulation

Learning Controllable Adaptive Simulation for Multi-resolution Physics

Tailin Wu* , Takashi Maruyama*, Qingqing Zhao*, Gordon Wetzstein, Jure Leskovec

ICLR 2023 notable top-25%
Best poster award at SUPR
We introduced the first fully deep learning-based surrogate models for physical simulations that jointly learn forward prediction and optimizes computational cost with RL.

ML for simulation

Artificial Intelligence for Science in Quantum, Atomistic, and Continuum Systems

Xuan Zhang, … Tailin Wu (31th/63 author), … Shuiwang Ji
Review paper summarizing the key challenges, current frontiers, and open questions of AI4Science, for quantum, atomistic, and continuum systems.

ML for simulation

Learning to Accelerate Partial Differential Equations via Latent Global Evolution

Tailin Wu, Takashi Maruyama, Jure Leskovec

NeurIPS 2022
We introduced a method for accelerating forward simulation and inverse optimization of PDEs via latent global evolution, achieving up to 15x speedup while achieving competitive accuracy w.r.t. SOTA models.

ML for simulation

Learning Large-scale Subsurface Simulations with a Hybrid Graph Network Simulator

Tailin Wu, Qinchen Wang, Yinan Zhang, Rex Ying, Kaidi Cao, Rok Sosič, Ridwan Jalali, Hassan Hamam, Marko Maucec, Jure Leskovec

SIGKDD 2022 & ICLR AI for Earth and Space Sciences Workshop long contributed talk
We introduced a hybrid GNN-based surrogate model for large-scale fluid simulation, with up to 18x speedup and scale to over 3D, 10^6 cells per time step (100x higher than prior models).

ML for discovery

ZeroC: A Neuro-Symbolic Model for Zero-shot Concept Recognition and Acquisition at Inference Time

Tailin Wu, Megan Tjandrasuwita, Zhengxuan Wu, Xuelin Yang, Kevin Liu, Rok Sosič, Jure Leskovec

NeurIPS 2022
We introduce a neuro-symbolic method that trained with simpler concepts and relations, can zero-shot generalize to more complex, hierarchical concepts, and transfer the knowledge across domains.

ML for simulation

Learning Efficient Hybrid Particle-continuum Representations of Non-equilibrium N-body Systems

Tailin Wu, Michael Sun, H.G. Jason Chou, Pranay Reddy Samala, Sithipont Cholsaipant, Sophia Kivelson, Jacqueline Yau, Zhitao Ying, E. Paulo Alves, Jure Leskovec†, Frederico Fiuza

NeurIPS 2022 AI4Science workshop, also under review
We introduced a hybrid particle-continuum representation for simulation of multi-scale, non-equilibrium, N-body physical systems, speeding up laser-plasma simulation by 8-fold with 6.8-fold error reduction.

Representation Learning

Graph Information Bottleneck

Tailin Wu, Hongyu Ren, Pan Li, Jure Leskovec

NeurIPS 2020
Featured in Synced AI Technology & Industry Review (机器之心)
We introduced Graph Information Bottleneck, a principle and representation learning method for learning minimum sufficient information from graph-structured data, significantly improving GNN’s robustness to adversarial attacks and random noise.

representation learning

Phase transitions for the Information Bottleneck in representation learning

Tailin Wu, Ian Fischer

ICLR 2020 We theoretically analyzed the Information Bottleneck objective, to understand and predict observed phase transitions (sudden jumps in accuracy) in the prediction vs. compression tradeoff.

ML for discovery

Toward an Artificial Intelligence Physicist for Unsupervised Learning

Tailin Wu, Max Tegmark
Physical Review E 100 (3), 033311

Featured in MIT Technology Review and Motherboard
Featured in PRE Spotlight on Machine Learning in Physics
We introduced a paradigm and algorithms for learning theories (small, interpretable models together with domain classifier) each specializing in explaining aspects of a dynamical system. It combines four inductive biases from physicists: divide-and-conquer, Occam’s razor with MDL, unification and lifelong learning.

ML for discovery

Discovering Nonlinear Relations with Minimum Predictive Information Regularization

Tailin Wu,Thomas Breuel, Michael Skuhersky, Jan Kautzin

Best Poster Award at ICML 2019 Time Series Workshop
We introduced a method for inferring Granger causal relations for large-scale, nonlinear time series with only observational data.

representation learning

Pareto-optimal data compression for binary classification tasks

Max Tegmark, Tailin Wu

Entropy 2020, 22(1), 7, as cover issue. arXiv:1908.08961. We introduce an algorithm for discovering the Pareto frontier of compression vs. prediction tradeoff in binary classification of neural networks.

representation learning

Learnability for the Information Bottleneck

Tailin Wu, Ian Fischer, Isaac Chuang, Max Tegmark

UAI 2019, Tel Aviv, Israel
Entropy 21(10), 924 (Extended version)
ICLR 2019 Learning with Limited Data workshop as spotlight
We theoretically derive the condition of learnability in the compression vs. prediction tradeoff in the Information Bottleneck objective.

ML for discovery

AI Feynman 2.0: Pareto-optimal symbolic regression exploiting graph modularity

Silviu-Marian Udrescu, Andrew Tan, Jiahai Feng, Orisvaldo Neto, Tailin Wu, Max Tegmark

NeurIPS 2020, Oral
We introduce a state-of-the-art symbolic regression algorithm that robustly re-discovering top 100 physics equations from noisy data from Feynman lectures.

representation learning

Learning with Confident Examples: Rank Pruning for Robust Classification with Noisy Labels

Curtis G. Northcutt*, Tailin Wu*, Isaac Chuang

UAI 2017 Cleanlab is built on top of it. We introduce a rank pruning method for classification with noisy labels, which provably obtains similar performance as without label noise.

ML for discovery

ViRel: Unsupervised Visual Relations Discovery with Graph-level Analogy

Daniel Zeng*, Tailin Wu*, Jure Leskovec

ICML 2022 Workshop of Beyond Bayes: Paths Towards Universal Reasoning Systems
We introduce a method that discovers common relational structures (analogical reasoning) from few-shot examples.

ML for discovery

Intelligence, physics and information – the tradeoff between accuracy and simplicity in machine learning

Tailin Wu

PhD Thesis
My PhD thesis summarizing my PhD work on AI for Physics and Physics for AI, together with representation learning with information theory.

ML for discovery

Toward a more accurate 3D atlas of C. elegans neurons

Michael Skuhersky, Tailin Wu, Eviatar Yemini, Amin Nejatbakhsh, Edward Boyden, Max Tegmark

BMC bioinformatics 23 (1), 1-18
We introduce a method for identifying the neuron ID in C. elegans and introduced a more accurate neuron atlas with the NeuroPAL technique.