Speaker: Zaid Harchaoui, NYU and Inria
Location: Warren Weaver Hall 1302
Date: February 27, 2015, 11:30 a.m.
Host: Subhash Khot
A typical feature of modern datasets is the large scale along several dimensions. This feature is pervasive in areas such as image categorization or video interpretation. Classical machine learning and statistical approaches cannot handle such large-scale datasets without further adaptation, both on the computational and the statistical sides. In this talk, we present several approaches tailored for computationally-efficient learning, that is learning with high predictive performance yet enjoying fast algorithmic convergence rate.
We introduce a family of first-order convex optimization algorithms for efficiently training algorithms. These algorithms rely on a latent atomic decomposition of the learning parameter, i.e. an expansion as a linear combination of elements in a possibly infinite dictionary. We present the generalized composite conditional gradient algorithm, and its variants and extensions to several learning settings: batch learning, online/stochastic learning, compressed learning, etc. We show theoretical convergence guarantees for this family of algorithms, which in particular highlight their near-optimality in high-dimensional settings. Such algorithms are successfully used for computationally efficient learning for collaborative filtering, learning embeddings of image categories, and visualization of heterogeneous large collection of images.
Zaid Harchaoui is currently Visiting Faculty in the Center for Data Science at the Courant Institute for Mathematical Sciences, as part of the Moore-Sloan Data Science Initiative. Zaid is on sabbatical leave from Inria, where is a tenured permanent researcher since 2010. He received his PhD from ParisTech (Paris, France), and his PhD thesis was on regularized kernel-based methods for detection. He received the Inria award for scientific excellence and the NIPS reviewer award. He gave a tutorial on "Frank-Wolfe, greedy algorithms, and friends" at ICML'14, and on "Large-scale visual recognition" at CVPR'13. He recently co-organized the workshop on “Optimization for Machine Learning” at NIPS'14, and the "Optimization and Statistical Learning" workshop in 2015 and 2013 in Ecole de Physique des Houches (France).
Refreshments will be offered starting 15 minutes prior to the scheduled start of the talk.