Deep Learning to Help People Understand Data and Understand Deep Networks

Large data sets are now generated by almost every activity in society, science, and commerce.

This generates an urgent need for methods for exploratory data analysis: How can a human analyst obtain an intuitive impression of the contents of a data set? This has been a key question in the closely intertwined research areas of statistics, machine learning, and data mining. Existing methods in these literature encompass a wide variety of methods, including dimensionality reduction methods, topic models, and association rule learning. At a very high level, all of these methods can be interpreted as search for a representation, such as a set of topic vectors or a list of association rules, that explains the data sufficiently well. The notion of “explaining the data” is quantified by measures such as marginal likelihood, confidence and lift (for rule mining), and so on, which can then be optimized algorithmically.

Although these methods have proved to be useful in practice, they suffer from the same drawback: the representations are chosen solely to explain the data and not directly to be interpretable by humans. A purported summary that is too complex, such as a topic model with 10,000 topics, or a list of 100,000 association rules, may itself be unsuited for human exploration and understanding, defeating the purpose of the method. This project aims to bridge this gap by developing new unsupervised machine learning methods aimed at helping people to understand large datasets and to be able to understand and interpret complicated deep networks.