Classification with unknown stochastic label noise: A distribution dependent view

Add to your list(s) Download to your calendar using vCal

• Dr Henry W J Reeve (Research Fellow, School of Computer Science, University of Birmingham, United Kingdom)
• Monday 29 July 2019, 11:00-12:00
• Computer Science, The Sloman Lounge (UG).

If you have a question about this talk, please contact Hector Basevi.

Host: Prof Ata Kaban

Abstract: The goal of supervised classification is to learn a classifier which takes as input a feature vector of properties characterising a particular object, and outputs a label which correctly assigns the object to a particular class. Examples range from assigning a piece of music to a particular genre based on an audio sample, through to predicting voting behaviour based on a set of demographic features. In order to learn a classifier, a supervised classification algorithm must have access to a set of training data consisting of labelled feature vectors. Typically, the theoretical analysis of classification algorithms makes the simplifying assumption that the training data have been correctly labelled. However, in a multitude of practical applications the labels within the training data have been perturbed by some unknown stochastic process. This is the challenge of classification with stochastic label noise.

Confronting the challenge of classification with stochastic label noise is an active research area within statistics and machine learning. In this talk we shall focus on the challenge of classification with unknown stochastic label noise, where the learner does not have access to any prior knowledge of the label noise. We shall discuss prior work in this domain which has established rates of convergence under strong separability conditions as well as a prior impossibility result that gives some grounds for scepticism about learning with unknown label noise. We shall then present some highlights from our work on learning with unknown label noise which builds upon recent developments in non-parametric statistics to give a distribution dependent view.

Firstly, we establish nearly optimal finite sample convergence rates for an empirically successful Robust k-NN algorithm due to Gao et al. (2018). We then go on to continue a more flexible setting with unbounded feature spaces in which we identify the optimal achievable convergence rates with a new adaptive algorithm and an information theoretic lower bound. This unbounded setting gives rise to a fascinating dichotomy in learning behaviour which hinges upon the rate at which class-conditional probabilities approach their extrema. Finally, we shall discuss a surprising connection between our results on classification with unknown stochastic label noise and results on randomised optimisation.

This is a joint work with Professor Ata Kaban.

Website: http://www.cs.bham.ac.uk/~reevehwj/

This talk is part of the Artificial Intelligence and Natural Computation seminars series.

This talk is included in these lists:

• Artificial Intelligence and Natural Computation seminars
• Computer Science Departmental Series
• Computer Science Distinguished Seminars
• Computer Science, The Sloman Lounge (UG)
• computer sience

Note that ex-directory lists are not shown.