Abstract:

Computer vision systems today exhibit a rich and accurate understanding of the visual world, but increasingly rely on learning on large labeled datasets to do so. This reliance on large labeled datasets is a problem especially when one considers difficult perception tasks, or novel domains where annotations might require effort or expertise. We thus need our machine vision systems to learn from very limited training data, a task at which we humans are adept. This is of course an ill-posed machine learning problem: how can one learn what is not in the data? The key is bringing to bear the right kind of inductive biases. In this talk I will present three results on learning from a few examples that explore different ways of introducing such biases: our initial steps towards building towards truly general-purpose perception systems that can learn in any new domain from very few training examples.

BIO:

Bharath Hariharan is an assistant professor in Cornell University. He was previously a postdoctoral researcher at Facebook AI Research. He received his PhD from University of California, Berkeley under the tutelage of Prof. Jitendra Malik. His research focuses on training computer vision systems with very limited training data.

Homepage:  http://home.bharathh.info/