[Oct. 2020]
  • Our three popular image enhancement algorithms: AOD-Net, EnlightenGAN, and DeblurGAN-V2, are included into the open-source GNU Image Manipulation Program toolbox (GIMP-ML), as deep-dehazing, enlighten, and deblur plugins.
[Sep. 2020]
  • 1 IEEE Trans. PAMI (privacy-preserving visual recognition) accepted
  • 8 NeurIPS'20 (learning to optimize + fast adversarial training + BERT lottery ticket + meta learning + robust contrastive learning + graph contrastive learning + ShiftAddNet + efficient quantized training) accepted
[Aug. 2020] [Jul. 2020]
  • 3 ECCV'20 (GAN compression + sketch-to-image synthesis + on-device learning-to-optimize) accepted
  • 1 ACM Multimedia (MMHand synthesizer) + 1 InterSpeech (AutoSpeech) accepted
[Jun. 2020]
  • 6 ICML'20 (AutoML domain generalization + noisy label training + self-supervised GCN + DNN optimization + GAN compression + NAS for Bayesian models) accepted
  • Our group co-organized the CVPR 2020 UG2+ Workshop and Prize Challenge
[May. 2020]
  • 1 IEEE Trans. PAMI (image enhancement for visual understanding) + 1 IEEE Trans. Mobile Computing (adaptive model compression) accepted
[Apr. 2020]
  • 2 CVPR'20 workshop (efficient triplet loss + fine-grained classification) accepted
[Mar. 2020]
  • 1 ISCA'20 (algorithm-hardware co-design to reduce data movement) accepted
[Feb. 2020]
  • 3 CVPR'20 (self-supervised adversarial robustness + fast GCN training + indoor scene reasoning [Oral]) + 1 IEEE Trans. Image Processing (visual understanding in poor-visibility environments) accepted
  • Our group will co-organize the IJCAI 2020 BOOM Workshop
[Jan. 2020]
  • 1 AISTATS'20 (CNN uncertainty quantification) + 1 IEEE Trans. CSVT (GAN data augmentation) accepted

Research Interests

[A] As Goals -- Enhancing Deep Learning Robustness, Efficiency, and Privacy

We seek to build deep learning solutions that are way beyond just data-driven accurate predictors. In our opinion, an ideal model shall at least: (1) be robust to perturbations and attacks (therefore trustworthy); (2) be efficient and hardware-friendly (for deployments in practical platforms); and (3) be designed to respect individual privacy and fairness.

[B] As Toolkits -- Automated Machine Learning (AutoML), and Learning-Augmented Optimization

We are enthusiastic about the rising field of AutoML, on both consolidating its theoretical underpinnings and broadening its practical applicability. State-of-the-art ML systems consist of complex pipelines, with choices of model architectures, algorithms and hyperparameters, as well as other configuration details to be tuned for optimal performance. They further often need to be co-designed with multiple goals and constraints. We consider AutoML to be a powerful tool and a central hub, in addressing those design challenges faster and better.

[C] As Applications -- Computer Vision and Interdisciplinary Problems

We are interested in a broad range of computer vision problems, ranging from low-level (e.g, image reconstruction, enhancement and synthesis) to high-level topics (e.g., recognition, segmentation, and vision for UAV/autonomous driving). We are also growingly interested in several interdisciplinary fields, such as biomedical informatics, geoscience, and IoT.

Prospective students