Semi-Supervised Learning in Computers and Humans

Abstract: Children learn with supervision, as well as without supervision by
observing the world around them. Occasionally, daddy might tell his daughter
that the animal walking by is called a "dog" but more frequently she will
encounter various animals without being told the names. Such unsupervised
experiences may help children establish the conceptual boundary between dogs
and non-dogs, and thus shape word learning. In general, human word- and
category-learning may be influenced by both supervised experiences (in which
both the item and its category label are directly provided) and by unsupervised
experiences (in which the item is encountered but no label is provided). In
machine learning this scenario is known as Semi-Supervised Learning, which has
been a fast-growing field of research.

This interdisciplinary talk explores semi-supervised learning in machine
learning and psychology: Do computers and humans learn from both supervised and
unsupervised experiences? We will first review the role of probabilistic
generative models (e.g. Gaussian mixture models) in semi-supervised machine
learning. We then present a human behavioral study in the form of shape
categorization. Our results showed that, although initial decision boundaries
were determined by the labeled examples, after exposure to the unlabeled
examples, human subjects shifted their boundaries. In this respect, the human
behavior conformed well to the predictions of a Gaussian mixture model for
semi-supervised learning. We conclude with differences between the human
behavior and machine learning model predictions, suggesting some fruitful
avenues for future inquiry.

Bio: Xiaojin Zhu is an Assistant Professor in Computer Sciences at University
of Wisconsin, Madison. His research interests include statistical machine
learning and applications to natural language tasks. Prof. Zhu received his PhD
in Language Technologies from Carnegie Mellon University in 2005, and BS and MS
degrees in Computer Science from Shanghai Jiao-Tong University, China.