Online Machine Learning for Streaming Scenarios

First 2012 Meeting of Research Programme in Machine Learning.

3.30 p.m. Tues 28 February 2012. G03 Inf. Forum<\p>

Agenda

3.30pm-4.30pm
Open Research Discussion. Various discussant will introduce particular issues in Online Machine Learning from their perspective, as a prompt for contribution from attendees.
4.30pm-5.00pm
Small group discussions (for those interested).
5.00pm-5.30pm
Discussion on current calls (for those interested).

Background

Historically, most machine learning methods have been developed in batch settings. The assumption is made that parameter estimation, structure learning, model comparison etc. is done with all the data already available. In addition the computational limitations are flexible: more computation simply takes more time.

In the modern world, this is an inappropriate way of thinking about a large number of data sources. Many data sources are streaming, and arrive in a continuous manner. In many signal processing and some machine learning scenarios there was no problem in doing inference online. However the reality is that methods need to be adaptable, transferable and flexible, and so online learning is also crucial.

Some common examples of situations where this occurs is in web based data, user click-through data, condition monitoring, human motion and robotics, large scale physics (particle physics, astronomy surveys etc.), speech and language systems, active vision, computer systems, system security, and many, many more. There are common challenges in all these scenarios: we need to handle data efficiently, we need to deal with time limited computation, we need to be flexible to system changes, and need models that perform better as more data is acquired without growing computational cost.

In this meeting we will examine problems and approaches in this space from a number of people?s perspectives. The meeting will not involve talks or posters or other fixed forms of presentation. Instead it will be an interactive meeting for all, with three parts.

First we will have a full session for everyone, where a number of people (the discussants) will be given the space to say something about that the major issues are for them, in their situation. We do not expect people to put in significant talk-like preparation for this, or else it might turn more into a talk, we simply ask those who will be given the floor to put in some brief consideration of what they might wish to mention. Other people can chip in, ask questions, or respond. We will discourage the use of presentation slides, but encourage the use of visual aids, whiteboard or flipchart. The session will be chaired.

Second, for the more interested parties, we will split up into small groups. Groups will be disbanded and reconfigured after very short periods of time to promote an online approach to ideas formation! This will give a chance for each person to see and reflect on a number of issues that others care about.

Finally we will introduce current funding calls related to this subject, and break into groups to discuss ideas around the funding call. Some groups might feel the chairs in a local pub provide a more comfortable place to discuss this that the open spaces of the Informatics Forum.

This page is maintained by Amos Storkey


Home : Research : Programmes : Machine-learning 

Informatics Forum, 10 Crichton Street, Edinburgh, EH8 9AB, Scotland, UK
Tel: +44 131 651 5661, Fax: +44 131 651 1426, E-mail: school-office@inf.ed.ac.uk
Please contact our webadmin with any comments or corrections. Logging and Cookies
Unless explicitly stated otherwise, all material is copyright © The University of Edinburgh