University Homepage School Homepage School Contacts School Search

 

Informatics Postgraduate Conference Talks

Paul Cronin

Virtual Reality Takes a Hike: an evaluation of a virtual Geology field course

David Sterratt

Is a biological temporal learning rule compatible with learning synfire chains?

Padraic Monaghan

Learning from graphical and sentential representations: explaining individual differences in computational terms

Herman M Gomes

Learning and Recognising Primal Sketch Features in a Log-polar Image Representation

Abstract

Traditional image feature extraction operators have usually been designed by hand, work independently of each other and act on Cartesian images (an artifact of sensor architecture). However, the architecture of the primate vision system seems to be quite different, and we can use this to produce interesting results in artificial vision systems.

In this talk I will present an account of the process used to build a log-polar image representation composed of low-level features extracted using a connectionist approach. The low level features (`edges', `bars', `blobs' and `ends') are based on Marr's primal sketch hypothesis for the human visual system and are used as the entry module of an iconic vision system.

This unusual but interesting image representation has been created using a neural network which learns examples of the features in a window of receptive fields of the image representation. An architecture designed to encode the feature's class, position, orientation and contrast has been proposed and tested. Success depended on the incorporation of a function that normalises the feature's orientation and a PCA pre-processing module to produce better separation in the feature space. A bootstrapping strategy that uses synthetic and real features has been used for the learning process.

Frank Keller

On the Origins of Gradient Acceptability

Abstract:

Experimental results consistently show that linguistic acceptability is a gradient concept. Native speakers' intuitions do not seem to support a binary distinction between acceptable and unacceptable linguistic structures: instead of a step function, we observe a smooth transition between the two extremes.

The present research aims to make sense of the phenomenon of gradient acceptability by associating different types of linguistic constraints with distinct acceptability effects. We present experimental results from a range of phenomena (extraction, gapping, complement order) that support a distinction between hard and soft linguistic constraints. Violations of hard constraints lead to strong unacceptability and fail to induce gradient behavior. Violations of soft constraints, on the other hand, cause only mild unacceptability, and violations differ in the degree of unacceptability they cause. Cumulativity effects can be observed for both types of constraints, which contributes to the gradient nature of linguistic acceptability.

>From a theoretical perspective, hard constraints seem to be structural in nature, while soft constraints seem to be associated with the interface between syntax and lexical semantics, discourse semantics, and information structure. We therefore expect linguistic context to interact with soft constraints, but not with hard ones, a prediction that is born out in our experimental data.

Jason Baldridge

"Incorporating sets into Combinatory Categorial Grammar without increasing its generative power."

Peter Hancock

"Proofs of wellfoundedness, programs, and their opposites"

Gann Bierner

"Using Discourse Clues in Information Retrieval"