Dapeng Hu

Hi! I am with Apple AIML. Previously, I was a scientist at A*STAR's Centre for Frontier AI Research (CFAR). I completed my Ph.D. at the National University of Singapore (NUS) in 2023, fortunately supervised by Jiashi Feng and Xinchao Wang. My research specializes in transfer learning and representation learning for computer vision. Before my Ph.D., I graduated with distinction in electronic engineering from Nanjing University.

Email  /  Google Scholar  /  LinkedIn

profile photo
Research

My research interests lie in generalizable and trustworthy deep learning with minimal human annotations. Specifically, I'm interested in transfer learning and representation learning, covering but not limited to the following topics:

  • How to assure alignment and safety of foundation models such as LLMs?
  • Test-time learning: How to ensure safe test-time model adaptation, selection, and calibration with access to only unlabeled testing data?
  • Pre-training and self-supervised learning: How to pre-train a generalizable foundation model with massive uncurated data?
  • Model-based transfer learning and fine-tuning: How to transfer the knowledge of a pre-trained source model or foundation model to a different downstream domain or task?
  • Large-scale empirical studies on deep learning models: How to uncover, understand, and explain the behavior of foundation models?
Towards Reliable Model Selection for Unsupervised Domain Adaptation: An Empirical Study and A Certified Baseline
Dapeng Hu, Mi Luo, Jian Liang†, Chuan-Sheng Foo
Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track, 2024.
A four-pages short version is in NeurIPS Workshop on Distribution Shifts, 2023.
arXiv /  code

We proposed EnsV, a highly simple, versatile, and reliable approach for practical model selection in unsupervised domain adaptation.
Pseudo-Calibration: Improving Predictive Uncertainty Estimation in Unsupervised Domain Adaptation
Dapeng Hu, Jian Liang, Xinchao Wang, Chuan-Sheng Foo†
International Conference on Machine Learning (ICML), 2024.
A four-pages short version is in NeurIPS Workshop on Distribution Shifts, 2023.
arXiv /  code

We proposed PseudoCal, a novel and versatile post-hoc framework for calibrating predictive uncertainty in unsupervised domain adaptation.
Mixed Samples as Probes for Unsupervised Model Selection in Domain Adaptation
Dapeng Hu, Jian Liang†, Jun Hao Liew, Chuhui Xue, Song Bai, Xinchao Wang†
Advances in Neural Information Processing Systems (NeurIPS), 2023.
arXiv /  code

We proposed MixVal, a novel target-only validation method for unsupervised domain adaptation with state-of-the-art model selection performance and improved stability.
UMAD: Universal Model Adaptation under Domain and Category Shift
Jian Liang*, Dapeng Hu*, Jiashi Feng, Ran He
Technical report
arXiv

We proposed a novel and effective method UMAD to tackle realistic open-set domain adaptation tasks where neither source data nor the prior about the label set overlap across domains is available for target domain adaptation.
DINE: Domain Adaptation from Single and Multiple Black-box Predictors
Jian Liang, Dapeng Hu, Jiashi Feng, Ran He
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. Oral
arXiv  /  code

We studied a realistic and challenging domain adaptation problem and proposed a safe and efficient adaptation framework (DINE) with only black-box predictors provided from source domains.
How Well Does Self-Supervised Pre-Training Perform with Streaming Data?
Dapeng Hu*, Shipeng Yan*, Qizhengqiu Lu, Lanqing Hong†, Hailin Hu, Yifan Zhang, Zhenguo Li, Xinchao Wang†, Jiashi Feng
International Conference on Learning Representations (ICLR), 2022.
arXiv

We conducted the first thorough empirical evaluation to investigate how well self-supervised learning (SSL) performs with various streaming data types and diverse downstream tasks.
Adversarial Domain Adaptation With Prototype-Based Normalized Output Conditioner
Dapeng Hu, Jian Liang†, Hanshu Yan, Qibin Hou, Yunpeng Chen
IEEE Transactions on Image Processing (TIP), Volume 30, 2021.
arXiv  /  code

We proposed a novel, efficient, and generic conditional domain adversarial training method NOUN to solve domain adaptation tasks for both image classification and segmentation.

Source Data-absent Unsupervised Domain Adaptation through Hypothesis Transfer and Labeling Transfer
Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, Jiashi Feng
IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2021.
arXiv  /  code

We enhanced SHOT to SHOT++ with an intra-domain labeling transfer strategy and achieved even better source data-free adaptation results than state-of-the-art data-dependent results.

No Fear of Heterogeneity: Classifier Calibration for Federated Learning with Non-IID Data
Mi Luo, Fei Chen, Dapeng Hu, Yifan Zhang, Jian Liang, Jiashi Feng
Advances in Neural Information Processing Systems (NeurIPS), 2021.
arXiv

We proposed CCVR, a simple and universal classifier calibration algorithm for federated learning.
Unleashing the Power of Contrastive Self-Supervised Visual Models via Contrast-Regularized Fine-Tuning
Yifan Zhang, Bryan Hooi, Dapeng Hu, Jian Liang, Jiashi Feng
Advances in Neural Information Processing Systems (NeurIPS), 2021.
arXiv  /  code

We proposed a theoretically and empirically promising Core-tuning method for fine-tuning contrastive self-supervised models.
Domain Adaptation with Auxiliary Target Domain-Oriented Classifier
Jian Liang, Dapeng Hu, Jiashi Feng
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021.
arXiv  /  code

We proposed ATDOC, a simple yet effective framework to combat classifier bias that provided a novel perspective addressing domain shift.
Do We Really Need to Access the Source Data? Source Hypothesis Transfer for Unsupervised Domain Adaptation
Jian Liang, Dapeng Hu, Jiashi Feng
International Conference on Machine Learning (ICML), 2020.
arXiv  /  code

We were among the first to look into a practical unsupervised domain adaptation setting called "source-free" DA and proposed a simple yet generic representation learning framework named SHOT.
A Balanced and Uncertainty-aware Approach for Partial Domain Adaptation
Jian Liang, Yunbo Wang, Dapeng Hu, Ran He, Jiashi Feng
European Conference on Computer Vision (ECCV), 2020.
arXiv  /  code

We tackled partial domain adaptation by augmenting the target domain and transforming it into an unsupervised domain adaptation problem
Professional Service

Journal Reviewer: TPAMI, IJCV, TIP, TMLR, TKDE

Conference Reviewer: ICML 2021-2023, NeurIPS 2021-2024, ICLR 2022-2025, CVPR 2022-2025, ICCV 2023, ECCV 2024, AISTATS 2025, AAAI 2025


Head TA, EE6934: Deep Learning (Advanced), 2020 Spring

Head TA, EE5934: Deep Learning, 2020 Spring

TA, EE4704: Image Processing and Analysis, 2019 Fall

TA, EE2028: Microcontroller Programming and Interfacing, 2019 Fall
Awards and Honors

Credit to Jon Barron.