Seminar

 
Abstract:
The recent success of deep convolutional neural networks (CNN) has made a breakthrough in visual object categorization, with the state-of-the-art model obtaining 80% classification accuracy on 1,000 object classes. While this is an impressive result, object categorization is yet to bring high impact to our everyday life, due to the small number of categories considered. There exist more than hundreds of thousands of nameable objects, and this set of categories is ever growing with the plethora of products that are newly introduced to our world every day. Thus, a truly practical categorization system should be able to recognize millions of object categories. However, the current state-of-the-art systems obtain only about 30% accuracy at maximum when classifying tens of thousands of classes, which limits their applicability to such large-scale cases. This low performance results from new challenges introduced in the large-scale recognition setting, such as finer granularity of categories, increased number of parameters and training time, class imbalance, and lack of training data. How should we tackle these challenges then? Fortunately, there exists a recognition system that is incredibly robust, scalable, and generalizes well – the human. In this talk, I will discuss about some models and algorithms I have developed to tackle new challenges posed by large-scale categorization, that are inspired by the human recognition and learning process.