I am a final-year Ph.D. student at the StarAI lab at University of California, Los Angeles, advised by Prof. Guy Van den Broeck. I am currently visiting Prof. Mathias Niepertβs lab at the University of Stuttgart.
π― Research highlights
My primary research focus is deep generative models (diffusion models [1,2,3], probabilistic circuits [5,6,7], variational autoencoders [4]). Other than understanding and mitigating the fundamental challenges toward good modeling performance [1,6,8], I am especially interested in efficient exact/approximate inference with guarantees of various deep generative models from both theoretical perspectives [9] and empirical perspectives [7,10].
π Research directions
-
What requirements must we impose on the structure of deep generative models to accurately and efficiently answer various probabilistic queries, such as computing arbitrary marginal probabilities or determining the MAP state? A useful theoretical framework for studying these problems is Probabilistic Circuits (PCs), which allows us to establish necessary and sufficient conditions on their structures to answer specific probabilistic queries [9]. I am among the first to significantly enhance the empirical performance of PCs, improving their effectiveness from struggling on MNIST to being compatible with variational autoencoders and even diffusion models on ImageNet32 [8]. To facilitate large-scale training and inference on PCs, I developed the Python package PyJuice, which is orders of magnitudes faster than all previous implementations.
-
Does PCs matter in the era of large language models? Many of my research works demonstrate that, with the ability to efficiently perform exact probabilistic inference, PCs can achieve better empirical performance on various down-stream tasks, either when used along [7] or when combined with other deep generative models [2,11].
-
How does the idea of tractable modeling generalizes to other types of deep generative models? In our recent work, we generalize the idea of combining the PC copula with a set of target univariate marginals to solve a fundamental problem that prevents discrete diffusion models from achieving strong performance with fewer steps β they fail to capture dependencies between output variables at each denoising step [1].
PyJuice
I am the main developer of PyJuice, which enables fast and scalable training and inference of Probabilistic Circuits. PyJuice has been used to train state-of-the-art PCs [8] and and has supported many related projects. Feel free to give it a try!
π₯ News
- 2024.10: Β ππ Our recent work on improving few-step generation performance of discrete diffusion models is now on ArXiv. Check it out at https://arxiv.org/pdf/2410.01949.
- 2024.09: Β ππ Check our recent work accepted to NeurIPS 2024. It demonstrates the importance of performing tractable inference in Offline Reinforcement Learning.
- 2024.08: Β ππ I gave an invitated talk about Probabilistic Circuits at Prof. Steffen Staabβs group at University of Stuttgart
- 2024.06: Β ππ I will co-organize the workshop on Open-World Agents at NeurIPS 2024
- 2024.05: Β ππ Our paper describing the technical details of PyJuice is accepted to ICML 2024.
π Educations
- 2020.09 - present, Computer Science Ph.D. student at UCLA, United States
- 2015.09 - 2019.06, Bachelorβs degree in Automation, Beihang University, China
π¬ Invited Talks
- 2024.08, Scaling Up Tractable Probabilistic Circuits for Inference-Demanding Application, University of Stuttgart, Germany
- 2023.07, Tractable Probabilistic Circuits, Dagstuhl, Germany
- 2023.05, Scaling Up Probabilistic Circuits by Latent Variable Distillation, ICLR oral presentation
- 2022.02, Tractable Probabilistic Circuits, Peking University, China
π Teaching
- Teaching assistant, Reinforcement Learning, University of Stuttgart, Spring 2024
- Lecturer (with Mathias Niepert), Introduction to AI, University of Stuttgart, Winter 2024
π Services
- PC member of ICML, NeurIPS, ICLR, AISTATS, AAAI
- Co-organizer of the Workshop on Open-World Agents at NeurIPS 2024
π Publications
Below is a list of selected publications. Please refer to my Google Scholar page for the full list of publications.
Discrete Copula Diffusion
Anji Liu, Oliver Broadrick, Mathias Niepert, Guy Van den Broeck
ArXiv / Paper
A Tractable Inference Perspective of Offline RL
Xuejie Liu*, Anji Liu*, Guy Van den Broeck, Yitao Liang
OmniJARVIS: Unified Vision-Language-Action Tokenization Enables Open-World Instruction Following Agents
Zihao Wang, Shaofei Cai, Zhancun Mu, Haowei Lin, Ceyao Zhang, Xuejie Liu, Qing Li, Anji Liu, Xiaojian Ma, Yitao Liang
NeurIPS 2024 / Paper / Website / Code
Describe, Explain, Plan and Select: Interactive Planning with Large Language Models Enables Open-World Multi-Task Agents
Zihao Wang, Shaofei Cai, Guanzhou Chen, Anji Liu, Xiaojian Ma, Yitao Liang
NeurIPS 2023 (Best paper award at the TEACH workshop at ICML 2023) / Paper / Code
Sparse probabilistic circuits via pruning and growing
Meihua Dang, Anji Liu, Guy Van den Broeck
NeurIPS 2022 (Oral; top 1.9%) / Paper / Code
Tractable and Expressive Generative Models of Genetic Variation Data
Meihua Dang, Anji Liu, Xinzhu Wei, Sriram Sankararaman, Guy Van den Broeck
RECOMB 2022 / Paper / Code
A Compositional Atlas of Tractable Circuit Operations for Probabilistic Inference
Antonio Vergari, YooJung Choi, Anji Liu, Stefano Teso, Guy Van den Broeck
NeurIPS 2021 (Oral; top 0.6%) / Paper / Code
Tractable regularization of probabilistic circuits
Anji Liu, Guy Van den Broeck
NeurIPS 2021 (Spotlight; top 3.7%) / Paper / Code
Off-Policy Deep Reinforcement Learning with Analogous Disentangled Exploration
Anji Liu, Yitao Liang, Guy Van den Broeck
AAMAS 2020 / Paper / Code