In the past 30 years, tremendous progress has been made in building effective classification models. Despite the success, we have to realize that, in major AI challenges, the key bottleneck is not the quality of classifiers but that of features. Since 2006,
learning high-level features using deep architectures has become a big wave of new learning paradigms. In recent two years, performance breakthrough was reported in both image and speech recognition tasks, indicating deep learning are not something ignorable.
In this talk, I will walk through the recent works and key building blocks, e.g., sparse coding, RBMs, auto-encoders, etc. and list the major research topics, including modeling and computational issues. In the end, I will discuss what might be interesting
topics for future research.