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 various perturbations, distribution shifts, and adversarial attacks; (2) be efficient for both inference and training, including resource-efficient, data-efficient, and label-efficient; and (3) be designed to respect individual privacy and fairness.
We are enthusiastic about AutoML, on both consolidating its theoretical underpinnings and broadening its practical applicability. State-of-the-art ML systems consist of complex pipelines, with various design choices. We consider AutoML to be a powerful tool and a central hub, in addressing those design challenges faster and better. Our recent work focuses on the data-driven discovery of model architectures (i.e., neural architecture search) and training algorithms (a.k.a. “learning to optimize”). We are meanwhile devoted to studying uprising ML models that show to be potentially “universal” workhorses, such as transformers and graph neural networks.
We are interested in a broad range of computer vision problems, ranging from low-level (e.g, image reconstruction, enhancement, and generation) 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, IoT and cyber-physical systems.