Xin Wang

Xin Wang

Ph.D. Candidate

EECS, UC Berkeley

xinw [at] eecs [.] berkeley [.] edu

2121 Berkeley Way West,
Berkeley, CA 94704

[CV] [Google Scholar] [Github] [LinkedIn] [Twitter]

I am Xin Wang, a Ph.D. student at UC Berkeley, working with Prof. Trevor Darrell and Prof. Joseph E. Gonzalez. I am part of the BAIR Lab, RISE Lab, and BDD Lab. Prior to UC Berkeley, I obtained my B.S. degree in Computer Science from Shanghai Jiao Tong University .

My research interest lies at the intersection of computer vision, machine learning and learning systems. I take a holistic view of intelligent systems, aiming to build vision systems that are adaptive, continual, efficient and can operate in real-time. I describe the premise of next-generation visual intelligent systems as solving the “last-mile delivery” problem of computer vision, which requires models capable of adapting to the challenging application scenarios with limited data, annotation, and computation.


As a Computer Vision researcher with a computer system emphasis, I work on a full stack solution to build visual intelligent systems from data collection and annotation, algorithmic designs of adaptive models, to real-time model serving. For model designs, I pioneer dynamic neural networks where the network parameters and structures are dynamically determined by various test-time information, such as the input image itself or task descriptions. Such test-time adaptation approach bridges the gap between training and inference, allowing the models to continuously learn and adapt to changing environments. For example, I propose SkipNet, where the execution paths are dynamically decided on a per input basis to reduce inference cost while maintaining the full model capacity and accuracy. Some other examples are constructing task-aware feature embeddings through parameter generation and test-time fine-tuning for few-shot image classification and object detection. I also explore various paradigms for learning with fewer data and labels during training, such as continual learning, heterogenous multi-task learning and weakly/semi-supervised learning.

Recent News
11/2020: Selected as one of the rising stars in EECS 2020 by UC Berkeley.
08/2020: Co-orgnized the Women in Computer Vision (WiCV) workshop at ECCV 2020.
07/2020: Presented our paper on Frustratingly Simple Few-Shot Object Detection at ICML 2020. We find a simple fine-tuning approach can outperform all the previous meta learning appraoches for few-shot object detection.
07/2020: Co-organized the second Human in the Loop Learning (HILL) workshop at ICML 2020.
06/2020: Gave a keynote talk on Motion Understanding via Heterogenous Multitask Learning at the fifth BMTT MOT Challenge workshop at CVPR 2020.
06/2020: Presented our work on BDD100K, the largeset driving video dataset with a new benchmark of 10 heterogenous perception tasks, at CVPR 2020.
06/2020: Presented our work on ShapeProp, which utlizes an iterative message passsing method to propagate the saliency regions for semi-supervised instance segmentation, at CVPR 2020.
02/2020: Gave an invited talk at Max Planck Institute for Informatics, Saarbrcken, Germany on Towards Human-Level Recognition and Generalization via Dynamic Representations.
University of California, Berkeley
Ph.D. in Computer Science
Aug 2015 - Present
Shanghai Jiao Tong University
Bachelor of Computer Science, IEEE Pilot Class
Sept 2011 - Jun 2015