Hello! I am a postdoctoral researcher at MIT CSAIL working with Stefanie Jegelka. I obtained my PhD in Applied Mathematics from Peking University (PKU) in 2023, advised by Yisen Wang, Jiansheng Yang, and Zhouchen Lin. Prior to that, I got my bachelor's degrees also from PKU math. My works received Best ML Paper Award of ECML-PKDD 2021 and Silver Best Paper Award of ICML 2021 AdvML workshop.
Currently I am mostly interested in developing theoretical understandings and principled designs of foundation models from the following key aspects:
- Self-supervised Learning (SSL). SSL is the driving engine of foundation models during the pretraining stage. I am mostly interested in theoretically understanding how existing SSL methods (contrastive, masked, autoregresive, etc) work and how to design better alternatives in principle.
- Adversarial Learning. Powerful LLMs need to be aligned to human purposes with guardrails to avoid being abused. I explore when existing alignment measures will fail (e.g., by adversarial attacks or jailbreaks), and how to systematically develop robust foundation models against adversarial and real-world distribution shifts.
- Neural architectures. I am interested in understanding the inherent mechanisms of backbone neural architectures, such as, Transformers and Graph Neural Networks.
Contact: yifei_w at mit.edu / Google Scholar / Github / X (Twitter)
News
- 2024.04. π₯ Anthropic proposed many-shot jailbreaking, which simply extends our In-context Attack (ICA) from 5 shots to 256 shots and jailbreaks most prominent LLMs (e.g., GPT3.5, GPT4, Claude2, Llama2-70B, Mistral)!
- 2024.04. π Our ICLR 2024 work Do Generated Data Always Help Contrastive Learning? is featured by Synced (in Chinese) with >10k reads.
- 2024.03. Camera ready and code of ICLR 2024 papers are out (see links below).
- 2024.03. Three working papers were acecpted at ICLR 2024 workshops, covering no-label backdoor attack (DPFM), robust evaluation for image generation (DPFM), and adversarial OOD benchmark (DMLR).
- 2024.02. π I was honored to receive Wenjun Wu Outstanding Ph.D. Dissertation Runner-Up Award (top 14 nation-wide) from CAAI. Wenjun Wu invented Wu's method for automatic theorem proving and pioneered AI research in China. Thanks everyone!
- 2024.01. π₯ Three papers about SSL methods were accepted by ICLR 2024, covering 1) non-negative contrastive learning, 2) the use of generated data for self-supervised learning, and 3) discrete tokenization for visual representation learning (Spotlight).
- 2023.12. I have joined MIT CSAIL as a postdoc.
- 2023.09. Five papers were accepted by NeurIPS 2023, aiming to provide better understandings of graph CL, identifiability of CL, adversarial examples, robust overfitting, and canonicalizability of graph eigenvectors. Paper and code are released.
- 2023.09. I will be serving as an Area Chair for ICLR 2024.
- 2023.05. Two SSL papers were accepted by ICML 2023, which analyze the roles of multi-modal supervision (e.g., image-text pairs in CLIP) and weak supervision (e.g., noisy labels) in self-supervised learning.
- 2023.02. One paper on fairness in adversarial training was accepted by CVPR 2023.
- 2023.01. Five papers about contrastive learning (CL) and graph learning were accepted by ICLR 2023, covering training dynamics, dimensional collapse, dynamic augmentation, CL-inspired normalization, and unbiased graph sampling.
Publications (* marks equal contribution)
- Non-negative Contrastive Learning ICLR 2024 2024 PDF | Code
- Do Generated Data Always Help Contrastive Learning? ICLR 2024 2024 PDF | Code
- On the Role of Discrete Tokenization in Visual Representation Learning ICLR 2024 (Spotlight) 2024 PDF | Code
- How to Craft Backdoors with Unlabeled Data Alone? ICLR 2024 Workshop on Navigating and Addressing Data Problems for Foundation Models (DPFM) 2024 PDF
- Robust Assessment of Image Generation with Virtual Classifiers ICLR 2024 Workshop on Navigating and Addressing Data Problems for Foundation Models (DPFM) 2024 PDF
- OODRobustBench: a benchmark and large-scale analysis of adversarial robustness under distribution shift ICLR 2024 Workshop on Data-centric Machine Learning Research (DMLR): Harnessing Momentum for Science 2024 PDF
- Balance, Imbalance, and Rebalance: Understanding Robust Overfitting from a Minimax Game Perspective NeurIPS 2023 2023 PDF | Code
- Adversarial Examples Are Not Real Features NeurIPS 2023 2023 PDF | Code
- Architecture Matters: Uncovering Implicit Mechanisms in Graph Contrastive Learning NeurIPS 2023 2023 PDF | Code
- Identifiable Contrastive Learning with Automatic Feature Importance Discovery NeurIPS 2023 2023 PDF | Code
- Laplacian Canonization: A Minimalist Approach to Sign and Basis Invariant Spectral Embedding NeurIPS 2023 2023 PDF | Code
- On the Generalization of Multi-modal Contrastive Learning ICML 2023 2023 PDF | Code
- Rethinking Weak Supervision in Helping Contrastive Representation Learning ICML 2023 2023 PDF
- CFA: Class-wise Calibrated Fair Adversarial Training CVPR 2023 2023 PDF | Code
- Equilibrium Image Denoising with Implicit Differentiation IEEE Transactions on Image Processing (TIP) 2023 PDF
- A Message Passing Perspective on Learning Dynamics of Contrastive Learning ICLR 2023 2023 PDF | Code | Slides | Blog
- Towards a Unified Theoretical Understanding of Non-contrastive Learning via Rank Differential Mechanism ICLR 2023 2023 PDF | Code
- Rethinking the Effect of Data Augmentation in Adversarial Contrastive Learning ICLR 2023 2023 PDF | Code
- ContraNorm: A Contrastive Learning Perspective on Oversmoothing and Beyond ICLR 2023 2023 PDF | Code
- Unbiased Stochastic Proximal Solver for Graph Neural Networks with Equilibrium States ICLR 2023 2023 PDF
- What Contrastive Learning Learns Beyond Class-wise Features? ICLR 2023 Workshop on Mathematical and Empirical Understanding of Foundation Models (ME-FoMo) 2023 PDF
- Rethinking the Necessity of Labels in Backdoor Defense ICLR 2023 Workshop on Backdoor Attacks and Defenses in Machine Learning (BANDS) 2023 PDF
- On the Connection between Invariant Learning and Adversarial Training for Out-of-Distribution Generalization AAAI 2023 (Oral) 2023 PDF
- How Mask Matters: Towards Theoretical Understandings of Masked Autoencoders NeurIPS 2022 (Spotlight, Top 5%) 2022 PDF | Code | Slides
- Improving Out-of-distribution Robustness by Adversarial Training with Structured Priors NeurIPS 2022 (Spotlight, Top 5%) 2022 PDF | Code | Slides
- When Adversarial Training Meets Vision Transformers: Recipes from Training to Architecture NeurIPS 2022 (Spotlight, Top 5%) 2022 PDF | Code
- Variational Energy-Based Models: A Probabilistic Framework for Contrastive Self-Supervised Learning NeurIPS 2022 SSL Workshop 2022 PDF
- AggNCE: Asymptotically Identifiable Contrastive Learning NeurIPS 2022 SSL Workshop (Oral) 2022 PDF
- Efficient and Scalable Implicit Graph Neural Networks with Virtual Equilibrium IEEE BigData 2022 (Long Talk) 2022 PDF
- Optimization-Induced Graph Implicit Nonlinear Diffusion ICML 2022 2022 PDF | Code
- G2CN: Graph Gaussian Convolution Networks with Concentrated Graph Filters ICML 2022 2022 PDF
- Chaos is a Ladder: A New Theoretical Understanding of Contrastive Learning via Augmentation Overlap ICLR 2022 2022 PDF | Code | Slides
- A Unified Contrastive Energy-based Model for Understanding the Generative Ability of Adversarial Training ICLR 2022 (π Silver Best Paper Award @ ICML 2021 AdvML workshop) 2022 PDF | Slides | Award
- Residual Relaxation for Multi-view Representation Learning NeurIPS 2021 2021 PDF | Slides | Blog
- Dissecting the Diffusion Process in Linear Graph Convolutional Networks NeurIPS 2021 2021 PDF | Code | Slides | Blog
- Reparameterized Sampling for Generative Adversarial Networks ECML-PKDD 2021 2021 (π Best ML Paper Award (1/685) & Invited to Machine Learning Journal) PDF | Code | Slides | Media | Talk | Award
- Train Once, and Decode as You Like COLING 2020 2020 PDF
- Non-negative Contrastive Learning ICLR 2024 2024 PDF | Code
- Do Generated Data Always Help Contrastive Learning? ICLR 2024 2024 PDF | Code
- On the Role of Discrete Tokenization in Visual Representation Learning ICLR 2024 (Spotlight) 2024 PDF | Code
- How to Craft Backdoors with Unlabeled Data Alone? ICLR 2024 Workshop on Navigating and Addressing Data Problems for Foundation Models (DPFM) 2024 PDF
- Robust Assessment of Image Generation with Virtual Classifiers ICLR 2024 Workshop on Navigating and Addressing Data Problems for Foundation Models (DPFM) 2024 PDF
- Architecture Matters: Uncovering Implicit Mechanisms in Graph Contrastive Learning NeurIPS 2023 2023 PDF | Code
- Identifiable Contrastive Learning with Automatic Feature Importance Discovery NeurIPS 2023 2023 PDF | Code
- On the Generalization of Multi-modal Contrastive Learning ICML 2023 2023 PDF | Code
- Rethinking Weak Supervision in Helping Contrastive Representation Learning ICML 2023 2023 PDF
- A Message Passing Perspective on Learning Dynamics of Contrastive Learning ICLR 2023 2023 PDF | Code | Slides | Blog
- Towards a Unified Theoretical Understanding of Non-contrastive Learning via Rank Differential Mechanism ICLR 2023 2023 PDF | Code
- Rethinking the Effect of Data Augmentation in Adversarial Contrastive Learning ICLR 2023 2023 PDF | Code
- ContraNorm: A Contrastive Learning Perspective on Oversmoothing and Beyond ICLR 2023 2023 PDF | Code
- What Contrastive Learning Learns Beyond Class-wise Features? ICLR 2023 Workshop on Mathematical and Empirical Understanding of Foundation Models (ME-FoMo) 2023 PDF
- Rethinking the Necessity of Labels in Backdoor Defense ICLR 2023 Workshop on Backdoor Attacks and Defenses in Machine Learning (BANDS) 2023 PDF
- How Mask Matters: Towards Theoretical Understandings of Masked Autoencoders NeurIPS 2022 (Spotlight, Top 5%) 2022 PDF | Code | Slides
- Variational Energy-Based Models: A Probabilistic Framework for Contrastive Self-Supervised Learning NeurIPS 2022 SSL Workshop 2022 PDF
- AggNCE: Asymptotically Identifiable Contrastive Learning NeurIPS 2022 SSL Workshop (Oral) 2022 PDF
- Chaos is a Ladder: A New Theoretical Understanding of Contrastive Learning via Augmentation Overlap ICLR 2022 2022 PDF | Code | Slides
- A Unified Contrastive Energy-based Model for Understanding the Generative Ability of Adversarial Training ICLR 2022 (π Silver Best Paper Award @ ICML 2021 AdvML workshop) 2022 PDF | Slides | Award
- Residual Relaxation for Multi-view Representation Learning NeurIPS 2021 2021 PDF | Slides | Blog
- Reparameterized Sampling for Generative Adversarial Networks ECML-PKDD 2021 2021 (π Best ML Paper Award (1/685). Invited to Machine Learning Journal) PDF | Code | Slides | Media | Talk | Award
- How to Craft Backdoors with Unlabeled Data Alone? ICLR 2024 Workshop on Navigating and Addressing Data Problems for Foundation Models (DPFM) 2024 PDF
- OODRobustBench: a benchmark and large-scale analysis of adversarial robustness under distribution shift ICLR 2024 Workshop on Data-centric Machine Learning Research (DMLR): Harnessing Momentum for Science 2024 PDF
- Adversarial Examples Are Not Real Features NeurIPS 2023 2023 PDF | Code
- Balance, Imbalance, and Rebalance: Understanding Robust Overfitting from a Minimax Game Perspective NeurIPS 2023 2023 PDF | Code
- CFA: Class-wise Calibrated Fair Adversarial Training CVPR 2023 2023 PDF | Code
- Rethinking the Effect of Data Augmentation in Adversarial Contrastive Learning ICLR 2023 2023 PDF
- Rethinking the Necessity of Labels in Backdoor Defense ICLR 2023 Workshop on Backdoor Attacks and Defenses in Machine Learning (BANDS) 2023 PDF
- On the Connection between Invariant Learning and Adversarial Training for Out-of-Distribution Generalization AAAI 2023 (Oral) 2023 PDF
- Improving Out-of-distribution Robustness by Adversarial Training with Structured Priors NeurIPS 2022 (Spotlight, Top 5%) 2022 PDF | Code | Slides
- When Adversarial Training Meets Vision Transformers: Recipes from Training to Architecture NeurIPS 2022 (Spotlight, Top 5%) 2022 PDF | Code
- A Unified Contrastive Energy-based Model for Understanding the Generative Ability of Adversarial Training ICLR 2022 (π Silver Best Paper Award @ ICML 2021 AdvML workshop) 2022 PDF | Slides | Award
- Laplacian Canonization: A Minimalist Approach to Sign and Basis Invariant Spectral Embedding NeurIPS 2023 2023 PDF | Code
- Architecture Matters: Uncovering Implicit Mechanisms in Graph Contrastive Learning NeurIPS 2023 2023 PDF | Code
- ContraNorm: A Contrastive Learning Perspective on Oversmoothing and Beyond ICLR 2023 2023 PDF
- Unbiased Stochastic Proximal Solver for Graph Neural Networks with Equilibrium States ICLR 2023 2023 PDF
- Efficient and Scalable Implicit Graph Neural Networks with Virtual Equilibrium IEEE BigData 2022 (Long Talk) 2022 PDF
- Optimization-Induced Graph Implicit Nonlinear Diffusion ICML 2022 2022 PDF | Code
- G2CN: Graph Gaussian Convolution Networks with Concentrated Graph Filters ICML 2022 2022 PDF
- Dissecting the Diffusion Process in Linear Graph Convolutional Networks NeurIPS 2021 2021 PDF | Code | Slides | Blog
Selected Awards
Excellent Graduate (top 1 per department), Beijing, 2023
Excellent Graduate, Peking University, 2023
Baidu Scholarship Runner-Up (top 20 nation-wide), Baidu Inc, 2022
National Scholarship (top 0.1% nation-wide), China, 2021, 2022
Principal Scholarship (top 1% university-wide), Peking University, 2022
Best ML Paper Award (1/685), ECML-PKDD, 2021
Silver Best Paper Award, ICML AdvML workshop, 2021
Professional Services
- ML Conferences: NeurIPS (2022, 2023), ICML (2022), AISTATS (2024), LoG (2023), ECML-PKDD (2022)
- Other conferences: CVPR (2023, 2024), ICCV (2023), ACL (2020, 2021)
- Journal: IEEE TPAMI, TMLR
- ICLR (2024)