In Bengio, S. & Bourlard, H., editors, *Machine Learning for Multimodal Interactions: First International Workshop, MLMI, Lecture Notes in Computer Science*, volume LNCS 3361, pages 159–172, 2004. Springer-Verlag.

Paper abstract bibtex

Paper abstract bibtex

Combining multiple information sources, typically from several data streams is a very promising approach, both in experiments and to some extend in various real-life applications. However, combining too many systems (base-experts) will also increase both hardware and computation costs. One way to selecting a subset of optimal base-experts out of $N$ is to carry out the experiments explicitly. There are $2^N-1$ possible combinations. In this paper, we propose an analytical solution to this task when weighted sum fusion mechanism is used. The proposed approach is at least valid in the domain of person authentication. It has a complexity that is additive between the number of examples and the number of possible combinations while the conventional approach, using brute-force experimenting, is multiplicative between these two terms. Hence, our approach will scale better with large fusion problems. Experiments on the BANCA multi-modal database verified our approach. While we will consider here fusion in the context of identity verification via biometrics, or simply biometric authentication, it can also have an important impact in meetings because this \it a priori information can assist in retrieving highlights in meeting analysis as in ``who said what''. Furthermore, automatic meeting analysis also requires many systems working together and involves possibly many audio-visual media streams. Development in fusion of identity verification will provide insights into how fusion in meetings can be done. The ability to predict fusion performance is another important step towards understanding the fusion problem.

@inproceedings{poh:2004:mlmi, author = {N. Poh and S. Bengio}, title = {Towards Predicting Optimal Subsets of Base Classifiers in Biometric Authentication Tasks}, booktitle = {Machine Learning for Multimodal Interactions: First International Workshop, {MLMI}, Lecture Notes in Computer Science}, editor = {S. Bengio and H. Bourlard}, pages = {159--172}, year = 2004, publisher = {Springer-Verlag}, volume = {LNCS 3361}, url = {publications/ps/poh_2004_mlmi.ps.gz}, pdf = {publications/pdf/poh_2004_mlmi.pdf}, djvu = {publications/djvu/poh_2004_mlmi.djvu}, idiap = {publications/pdf/rr04-17.pdf}, original = {2004/optimal_subset_mlmi}, topics = {biometric_authentication}, web = {http://www.springerlink.com/index/Y2AN6DEWGKK26JPW}, abstract = {Combining multiple information sources, typically from several data streams is a very promising approach, both in experiments and to some extend in various real-life applications. However, combining too many systems (base-experts) will also increase both hardware and computation costs. One way to selecting a subset of optimal base-experts out of $N$ is to carry out the experiments explicitly. There are $2^N-1$ possible combinations. In this paper, we propose an analytical solution to this task when weighted sum fusion mechanism is used. The proposed approach is at least valid in the domain of person authentication. It has a complexity that is additive between the number of examples and the number of possible combinations while the conventional approach, using brute-force experimenting, is multiplicative between these two terms. Hence, our approach will scale better with large fusion problems. Experiments on the BANCA multi-modal database verified our approach. While we will consider here fusion in the context of identity verification via biometrics, or simply biometric authentication, it can also have an important impact in meetings because this {\it a priori} information can assist in retrieving highlights in meeting analysis as in ``who said what''. Furthermore, automatic meeting analysis also requires many systems working together and involves possibly many audio-visual media streams. Development in fusion of identity verification will provide insights into how fusion in meetings can be done. The ability to predict fusion performance is another important step towards understanding the fusion problem.}, categorie = {C}, }

Downloads: 0