Seminar: Variance inflation in high dimensional learning

Speaker: Lars Kai Hansen, Dept of Applied Mathematics and Computer Science, Technical University of Denmark
UCL Contact: Ingemar Cox (Visitors from outside UCL please email in advance).
Date/Time: 07 May 15, 11:00 - 12:00
Venue: 6.12

Abstract

Many important machine learning models are based on Euclidean distance or linear projections in high dimensional feature spaces. When adapting such models to small training sets we face the problem that the span of the training data is not the full input space. Hence, when models are applied to test data they are effectively blind to the missed orthogonal subspace. This shows up as an inflated variance of hidden variables in the training set and in general that training and test latent variables follow different probability laws. We will discuss basic means to detect and correct variance inflation for both unsupervised and supervised learning.

References:
T.J. Abrahamsen, L.K. Hansen. A Cure for Variance Inflation in High Dimensional Kernel Principal Component Analysis. Journal of Machine Learning Research 12:2027-44 (2011).

T.J. Abrahamsen, L.K. Hansen. Variance Inflation in High Dimensional Support Vector Machines. Pattern Recognition Letters 34(16): 2173-80.(2013).