EE Faculty Colloquium on Learning from unreliable data

Speaker: Prof. P.S. Sastry, Dept of Electrical Engineering, Indian Institute of Science

Abstract:

Supervised learning of classifiers is widely used in many applications of AI/ML. The deep networks used in such applications today need a large training set. Creating a labelled data set where one can have high confidence in the labels is both expensive and time consuming. Data sets created through crowd sourcing or automatic labelling methods normally have many random labelling errors. There is considerable amount of empirical evidence to show that standard algorithms are likely to do poorly when there is significant amount of label noise in the data. Hence it is interesting to ask whether one can design classifier learning algorithms that are robust to different types of random labelling errors (in the training data). Over the years this problem has been investigated by many researchers and many interesting ideas and algorithms for such robust learning are proposed. In this talk we present an overview of the problem of learning with noisily labelled training set and review some of the approaches proposed for tackling the problem. We concentrate mainly on risk minimization schemes. We discuss what are called symmetric loss functions and their role in robust risk minimization. We will also briefly discuss approaches based on sample selection and weighted risk minimization and present a sample selection algorithm based on batch statistics. The discussion would be biased towards some work done in our lab.

Speaker’s Bio:

P.S. Sastry obtained BSc in Physics from IIT, Kharagpur, and BE from ECE dept and PhD from EE dept at IISc. He has been a faculty member of dept EE, IISc, for more than 35 years now. His research interests include Pattern Recognition, Machine Learning, Data Mining, and Computational Neuroscience.