Building intelligent system that can emulate human reasoning process has been an ultimate goal of many philosophers, scientists, and researchers. When artificial intelligence emerged as a field of computer science, people thought that logic based systems may model the human reasoning process to a large extent. However, it turns out that human sensory systems are robust to uncertainty and noise of the external environment. Consequently, researchers have been considering probabilistic models in pursuit of building useful intelligent systems. It is this intersection where the goal of computer scientists and statisticians meets. In computer science, these focused studies are compiled to a new subject called Machine Learning.

In this talk, some recent probabilistic models, their applications, and limitations will be discussed with the goal of learning useful features. The motivation of feature learning stems from the research outcomes of neuroscientists. For example, in the vision, edges in a scene constitute the primary features for higher cognitive processes. These primary features in turn combine to form more complex features in a hierarchical manner. A class of probabilistic models called deep belief nets has been proved as the state of-the-art method for learning feature hierarchically. Several applications, including speech recognition, object and motion recognition, and natural language processing will be discussed in this talk. Although empirical successes prove the modeling strengths of deep belief nets, theoretical investigations and analysis are less explored. Furthermore, it is important to develop new algorithms that can learn accurate distribution of the data faster. One of my goals of this talk is to discuss the possibility of carrying out joint research with the ISRT, which is the leading statistical research institute in Bangladesh.