next up previous contents
Next: Task 5.4: Technical Description Up: Task 5.4: Incorporating Auditory Previous: Task 5.4: Task Objective

Task 5.4: Status

At FPMs, tests on non-linear dimensionality reduction algorithms have been performed.

At Sheffield we have concentrated on the development of automatic, unsupervised dimension reduction algorithms. We have adopted a latent variable modelling framework, in which it is assumed that the high dimensional observed data is generated by from a lower dimensional latent (or hidden) space. To avoid the additional complexities that would occur in developing these algorithms using auditory data, we have concentrated on an articulatory modelling problem (electropalatography). Dimension reduction algorithms investigated include factor analysis, mixtures of factor analysers, mixtures of Bernoulli and the generative topographic map (GTM).

Christophe Ris