Event

Multi-View Learning in the Presence of View Disagreement

Mario Christoudias

ICSI

Tuesday, October 28, 2008
12:30

Traditional multi-view learning approaches suffer in the presence of view disagreement, i.e., when samples in each view do not belong to the same class due to view corruption, occlusion or other noise processes. In this talk I will present a multi-view learning approach that uses a conditional entropy criterion to detect view disagreement. Once detected, samples with view disagreement are filtered and standard multi-view learning methods can be successfully applied to the remaining samples. Experimental evaluation on synthetic and audio-visual databases demonstrates that the detection and filtering of view disagreement considerably increases the performance of traditional multi-view learning approaches.

I will also briefly discuss work done in our group in collaboration with Stanley Peter's group at Stanford on multi-modal disambiguation of 'you' references in multi-party dialogues. Reference resolution is an important part of multi-party dialogue understanding. Previous approaches to you disambiguation have mainly focused on spoken language only. In this work we investigated whether visual cues such as head gaze and gesture can aid in resolving you references. Experiments on the AMI meeting corpus demonstrate that the use of visual cues gives a significant improvement over speech-only performance.

This work was done in collaboration with Raquel Urtasun and Trevor Darrell.