Date: Thursday 7.6.2018 16:15-17:00
Venue: room T5, CS building, Konemiehentie 2

Speaker:
Rohit Babbar

Title:
Algorithms for Extreme Multi-label Classification

Abstract:

In the era of big data, large-scale classification involving tens of thousand target categories is not uncommon. Also referred to as Extreme Classification, it has also been recently shown that the machine learning challenges arising in ranking, recommendation systems and web-advertising can be effectively addressed by reducing it to extreme multi-label classification framework. In this talk, I will discuss two of my recent works, targetting two different goals in extreme multi-label classification. Though conceptually simple, these methods outperform state-of-the-art methods, including those which are used for ad-prediction by Microsoft in production system of Bing Search. The training process for these algorithms makes use of openMP based distributed architectures, and is able to leverage thousands of cores for computation. I will also draw some connections to works which are aimed to designing efficient adversarial attacks and building defences for state-of-the-art deep learning systems.

**

Helsinki Algorithms Seminar is a weekly meeting of researchers in the Helsinki area interested in the art of algorithms and algorithm design, broadly interpreted to cover both theoretical ideas and algorithm engineering on concrete computing platforms. In most cases we have a presentation prepared for each meeting to communicate an idea, a recent result, work-in-progress, or demo, but this should not be at the expense of discussion and simply having fun with algorithms.

Our affiliations are with Aalto University and the University of Helsinki, and accordingly our activities alternate between the Otaniemi Campus of Aalto University and the Kumpula Campus of University of Helsinki, catalyzed by the Helsinki Institute for Information Technology HIIT, under the Algorithmic Data Analysis (ADA) programme.

For the season programme, please see the seminar webpage.