HMM topology optimization with anti-likelihood criterion for handwriting recognition

This paper investigates the effect of applying anti-likelihood into model selection criteria to optimizing the architecture of a continuous density HMM aimed at on-line handwriting recognition. In this paper, we proposed a systematic HMM topology optimization method that is able to improve discrimination power as well. We also compared our method with several existing optimization criteria including ML(Maximum likelihood) criterion and BIC. We also explored the relationship of the number of parameters and recognition ratio according to model selection criteria. We used anti-likelihood, which is defined as the probability of out of class data for given model, as well as likelihood to improve the discrimination power. Experimental results on the online handwritten character recognition task show that discriminative model selection approach exhibits the best performance with the less number of parameters.