Research Interests
Learning from Imperfect Data
Today, big data is an important key to the improvement of machine learning. However, annotating all the data is virtually impossible, considering the huge cost. One of the promising solutions is to utilize an algorithm that can learn with minimal supervision. To this end, we study visual intelligence methods that can understand the world with minimal (semi-, weakly-supervised learning) or no human supervision (self-supervised learning).
Reliable Machine Learning
Currently, state-of-the-art visual recognition systems are known to be vulnerable and hard to explain. This hinders the reliability of machine learning methods. To address these limitations, we are interested in building a visual recognition machine that is robust to (1) noise or attack, (2) dataset bias, and (3) data distribution shifts. In addition, we study how to correctly estimate the reasons for the decision made by a visual recognition method.
|
