Evaluation of incremental learning algorithms for an HMM-based handwritten isolated digits recognizer
Paulo Cavalin, Robert Sabourin, Ching Y. Suen, Alceu S. Britto · 2008
We present an evaluation of incremental learning algorithms for the estimate of HMM parameters. The main goal was to investigate incremental learning algorithms that can replace traditional batch learning techniques, incorporating the advantages of incremental techniques for designing complex pattern recognition systems. Experiments were carried out on isolated digits, extracted from the NIST SD19, by using a state-of-the-art HMM-based isolated digits recognizer. The experiments demonstrated that batch learning performs slightly better for generating classifiers with good generalization performance. However, the results obtained by the Ensemble Training algorithm are very encouraging for pursuing further research in this subject, given that the loss in terms of performance is relatively small. Furthermore, we demonstrated that the incremental approaches provide lower-cost algorithms, which is a valuable advantage.