Wei Hu
Assistant Professor, EECS – Computer Science and Engineering
Studies the theory and science of deep learning, including generalization, training dynamics and the behavior of large language models.
Wei Hu's lab at Michigan works on learning theory & optimization and large language models. 16 papers since 2023, 15 at top venues, mostly ICLR and NeurIPS. 2 PhD students. 1 alumnus. Applications close .
At a glance
- papers, 2023–2026
- 4/ year
- at T1 venues
- 94%
- PhD students
- 2
- MS students
- 0
- alumni listed
- 1
Papers by yearfrom the lab homepage, 2023–2026
- 2 papers2026 · T1 2 · collecting
- 5 papers2025 · T1 5
- 6 papers2024 · T1 5
- 3 papers2023 · T1 3
T1 venuesOtherVenues: ICLR, NeurIPS, TMLR, ACL, ICML
Research focuspaper titles that use each topic's words
- Learning theory & optimization6 papers
- generalization1 paper
- dynamics1 paper
- Large language models0 papers
- behavior0 papers
Paper list
From the lab homepage.
20262 papers
20255 papers
- NeurIPS (T1)Benign Overfitting in Single-Head Attention (opens in a new tab)
- ICLR (T1)Let Me Grok for You: Accelerating Grokking via Embedding Transfer from a Weaker Model (opens in a new tab)
- ICLR (T1)Swing-by Dynamics in Concept Learning and Compositional Generalization (opens in a new tab)
- JMLR (T1)Understanding Deep Representation Learning via Layerwise Feature Compression and Discrimination (opens in a new tab)
- NeurIPS (T1)What Happens During the Loss Plateau? Understanding Abrupt Learning in Transformers (opens in a new tab)
20246 papers
- NeurIPS (T1)Abrupt Learning in Transformers: A Case Study on Matrix Completion (opens in a new tab)
- —AISTATS 2024
- ICLR (T1)Benign Overfitting and Grokking in ReLU Networks for XOR Cluster Data (opens in a new tab)
- TMLR (T1)Bias Amplification Enhances Minority Group Performance (opens in a new tab)
- ICLR (T1)Dichotomy of Early and Late Phase Implicit Biases Can Provably Induce Grokking (opens in a new tab)
- ICLR (T1)How Do Transformers Learn In-Context Beyond Simple Functions? A Case Study on Learning with Representations (opens in a new tab)
20233 papers
- ICML (T1)Are Neurons Actually Collapsed? On the Fine-Grained Structure in Neural Representations (opens in a new tab)
- NeurIPS (T1)Going Beyond Linear Mode Connectivity: The Layerwise Linear Feature Connectivity (opens in a new tab)
- ICLR (T1)Implicit Bias in Leaky ReLU Networks Trained on High-Dimensional Data (opens in a new tab)
Where alumni went1 listed, 0% to industry
Industry 0Academia 1
Also to
- Harvard University
Apply to MichiganFall 2027
Michigan CSE Computer Science PhD
- Round
- Single round
- Opens
- —
- GRE
- See program page
- English
- See program page
- Fee
- See program page
Not yet confirmed for this intake.
Similar labsnearest by research summary
- USNYU
Rachel Greenstadt
- Trustworthy & safe ML
- USGeorgia Tech
Kartik Goyal
- Probabilistic & causal inference
- CNZhejiang
Siliang Tang
- Multimodal & grounded language
- CNSJTU
Yibo Yang
- Trustworthy & safe ML
- CNSJTU
Chaoxiang He
- Trustworthy & safe ML
- SGNTU
Boyang Li
- Multimodal & grounded language
Your notes
Save this lab to keep notes.
Data
Checked against the lab homepage.
- Collected
- Verified
- Papers
- The lab homepage
- Members
- Lab homepage
- Alumni
- Lab homepage
Tiers are venue classifications, not lab rankings.
Michigancloses