I am Xin Wang, a forth-year Ph.D. candidate in Computer Science at UC Berkeley, advised by Prof. Joseph E. Gonzalez and Prof.Trevor Darrell. I am part of the RISE Lab, BAIR Lab and BDD Lab.
Prior to UC Berkeley, I obtained my B.S. degree in Computer Science from Shanghai Jiao Tong University in 2015.
My research interest is in computer vision and learning system designs. My recent works focus on the design of dynamic neural networks for better generalization and inference efficiency in computer vision applications. Previously, I was also involved in the Clipper project for large-scale low-lentency machine learning model serving.
TAFE-Net: Task-Aware Feature Embeddings for Low Shot Learning
Conference on Computer Vision and Pattern Recognition (CVPR) 2019
Novel task-aware feature embeddings for zero-shot and few-shot learning
Accel: A Corrective Fusion Network for Efficient Semantic Segmentation on Video
Conference on Computer Vision and Pattern Recognition (CVPR) 2019, Oral presentation
Efficient semantic video segmentation system by network fusion
Deep Mixture of Experts via Shallow Embedding
Conference on Uncertainty in Artificial Intelligence (UAI) 2019
Channel level dynamic routing for efficient inference
SkipNet: Learning Dynamic Routing in Convolutional Networks
European Conference on Computer Vision (ECCV) 2018
Dynamically bypassing convolutional layers on a per-input basis for effcient prediction
IDK Cascades: Fast Deep Learning by Learning not to Overthink
Conference on Uncertainty in Artificial Intelligence (UAI) 2018
Dynamically compose models with different accuracies and computational costs for efficient prediction
Clipper: A Low-Latency Online Prediction Serving System
USENIX Symposium on Networked Systems Design and Implementation (NSDI) 2017
Clipper is a prediction serving system that sits between user-facing applications and a wide range of commonly used machine learning models and frameworks
Scalable Training and Serving of Personalized Models
A MTL formulation to simultaneously address both the challenge of changing data and enable low-latency highly available model serving.