School of Informatics, University of Edinburgh, UK
Wednesday 16 May 2007
Seminar Room B10 (Basement)
Alexandra House, 17 Queen Square, London, WC1N 3AR
Coupled HMMs for Primitive-Timing Representations of Handwriting
Biological movement is built up of subblocks or motor primitives. Hence we would also expect handwriting to be decomposable into primitives. We show how a Factorial HMM based model encoding what we call a piano model can learn appropriate primitives from unpartitioned handwriting data. The posterior probability of primitive activation is abstracted into a spike timing representation of the data, which we model using a separate HMM. This complete hierarchical model can be used to generate samples capturing the variation of the dataset, based upon a compact spike timing code. Furthermore, the timing code dictates which character is being reproduced, without which, a primitive babbling produces a scribbling style output which captures an aspect of the dataset, but displays no global character coherence.
This work is joint work with Ben Williams and Marc Toussaint.