17 May 2006

2.30pm

Gatsby Unit Annual Seminar

We are delighted to announce the first in an annual series of Gatsby Seminars, with talks by distinguished researchers in theoretical neuroscience and machine learning.

This year's talks will be given by Dr Li Zhaoping, from the Dept of Psychology at UCL, and Prof John Shawe-Taylor, from the Dept of Computer Science at the University of Southampton. They will be held at 2.30pm in the Leolin Price Lecture Theatre, Institute of Child Health, 30 Guilford Street, London WCIN 1EH and will be followed by a wine reception at 5.00pm at The Gatsby Unit, Alexandra House, 17 Queen Square

Titles and abstracts below.

All are welcome.

Dr Li Zhaoping                                                                                                                     2.30pm

The More Briefly one Looks, the More Effectively one `Sees': Vision by V1

I show that a visual search task can be better performed when one views the search array for a shorter time, and suggest an account of this phenomenon based on an analysis of V1's contribution to vision. The cost of a prolonged view comes from the interference of higher level object recognition on lower level image feature processing. A similar effect underlies the trick for art novices of drawing a portrait upside down in order to reproduce lower level image features, such as contours, with less interference from higher level face cognition.

In our task, the search target has an uniquely oriented bar but is identical in shape to distractors. Lower level image feature processes enable the unique orientation to pop out, attracting gaze towards the target. Subsequently, higher level object processes, involving focused attention, recognize the target object in a viewpoint invariant manner, confusing the target as being a distractor and interfering with the task.  Lower and higher processes lead to their respective behavioural decisions manifested in eye movements and ultimate task performances.

I will show physiological and computational evidence implicating V1 mechanisms for the lower level feature pop out, and review data about higher object processes in higher brain areas.

Prof John Shawe-Taylor                                                                                                      3.45pm

Inferring Semantic Representations from Data

The talk addresses the question of how effectively we can learn underlying semantics from data. We concentrate on text analysis as a domain where semantics are relatively cleanly defined and on which learning approaches have made significant advances. The links between Latent Semantic Indexing, Latent Semantic Kernels and kernel Principal Components Analysis are discussed and the generalisation of such representations is discussed.  Cross-lingual information retrieval suggests the use of Canonical Correlation Analysis as a Semantic inference tool. Again a kernel version can be defined and with appropriate regularisation applied in high-dimensional feature spaces. Applications of the same approach to non-text data will also be presented.