Robust Low-Dimensional Structure Learning for Big Data and its Applications

Jiashi Feng

Tuesday, July 1
12:30 p.m., Conference Room 5A

The explosive growth of data in the era of big data has presented great challenges to traditional structure learning techniques. In this thesis, we propose deterministic and online learning methods for robustly recovering the low-dimensional structure of big data.

We first develop a DHRPCA method to recover low-dimensional subspace of high-dimensional data. DHRPCA possesses maximal robustness, and is asymptotically consistent in the high-dimensional space. Moreover, it exhibits significantly high efficiency for handling big data.

Secondly, we propose two online learning methods, OR-PCA and online RPCA, to further enhance the scalability for robustly learning the low-dimensional structure of big data, under limited memory and computational cost budget.

Thirdly, we devise two low-dimensional learning algorithms for visual data analysis: (1) geometric feature pooling, and (2) auto-grouped sparse representation. These two methods achieve state-of-the-art performance on several benchmark image classification datasets.

Bio:

Jiashi Feng receives his phd degree from National University of Singapore (NUS) in 2014. Now he works as a postdoc with Prof. Trevor Darrell. He got his bachelor degree in automation from University of Science and Technology of China (USTC). He is interested in both computer vision and machine learning. In particular, his research works focus on object recognition, attributes learning, robust optimization, online and distributed learning. He received the best technical demo award in Multimedia 2012.