Classification with HMMs

A CRF is a discriminative probabilistic model. Van Kasteren et al. [32] compared CRF with HMMs and found that CRF outperformed HMMs in all cases with respect to time slice accuracy, but HMMs achieved the overall highest accuracy. This is due to the way both models maximize their parameters. HMMs make use of a Bayesian framework in which a separate model is learned for each class. A CRF uses a single model for all classes. A comparison of HMMs and CRF was also discussed by Hu et al. [33], who found that CRF is able to easily incorporate a wide variety of computed features, which allows domain knowledge to be added to the models. They also showed that due to the independence assumptions inherent in HMMs, such computed features are not nearly as effective in improving classification accuracy. Thus, CRF s classification accuracy has shown to be consistently higher than HMM s. 

Figure 7.1. The trend analysis strategy using HMMs. The process information (measurement) at a time instant k is first expressed as a fuzzified sequence (FS) and then processed through a classification step. Reprinted from [340]. Copyright 1998 with permission from Elsevier.

SUPERFAMILY [19] is a collection of profile HMMs aiming to represent all proteins of known structure. Each model corresponds to a domain described in the SCOP structural classification database and aims to describe the entire SCOP superfamily associated with the domain. 

While many of these points are valid, there are often solutions which help alleviate any problems. As just explained, the use of d5mamic features helps greatly with the problems of observation independence and discrete states. As we shall see the linearity issue is potentially more of a problem in speech s mthesis. Models such as neural networks which perform classification directly have been proposed [375] and have produced reasonable results. More recently, discriminative training has become the norm in ASR [495], [360] where HMMs as described are used, but where their parameters are trained to maximise discrimination, not data likelihood. 

---