Adaptive systems for hidden Markov model-based pattern recognition systems.

摘要

This thesis focuses on the design of adaptive systems (AS) for dealing with complex pattern recognition problems. Pattern recognition systems usually rely on static knowledge to define a configuration to be used during their entire lifespan. However, some systems need to adapt to knowledge that may not have been available in the design phase. For this reason, AS are designed to tailor a baseline pattern recognition system as required, and in an automated fashion, in both the learning and generalization phases. These AS are defined here, using hidden Markov model (HMM)-based classifiers as a case study.;We first evaluate incremental learning algorithms for the estimation of HMM parameters. The main goal is to find incremental learning algorithms that perform as well as the traditional batch learning techniques, but incorporate the advantages of incremental learning for designing complex pattern recognition systems. Experiments on handwritten characters have shown that a proposed variant of the Ensemble Training algorithm, which employs ensembles of HMMs, can lead to very promising results. Furthermore, the use of a validation dataset demonstrates that it is possible to achieve better performances than those of batch learning.;We then propose a new approach for the dynamic selection of ensembles of classifiers. Based on the concept called "multistage organizations", the main objective of which is to define a multi-layer fusion function that adapts to individual recognition problems, we propose dynamic multistage organization (DMO), which defines the best multistage structure for each test sample. By extending Dos Santos et al's approach, we propose two implementations for DMO, namely DSAm and DSAc. DSAm considers a set of dynamic selection functions to generalize a DMO structure, and DSAc uses contextual information, represented by the output profiles computed from the validation dataset. The experimental evaluation, considering both small and large datasets, demonstrates that DSAc outperforms DSAm on most problems. This shows that the use of contextual information can result in better performance than other methods. The performance of DSAc can also be enhanced in incremental learning. However, the most important observation, supported by additional experiments, is that dynamic selection is generally preferred over static approaches when the recognition problem presents a high level of uncertainty.;Finally, we propose the LoGID (Local and Global Incremental Learning for Dynamic Selection) framework, the main goal of which is to adapt hidden Markov model-based pattern recognition systems in both the learning and generalization phases. Given that the baseline system is composed of a pool of base classifiers, adaptation during generalization is conducted by dynamically selecting the best members of this pool to recognize each test sample. Dynam selection is performed by the proposed K-nearest output profiles algorithm, while adaptation during learning consists of gradually updating the knowledge embedded in the base classifiers by processing previously unobserved data. This phase employs two types of ncremental learning: local and global. Local incremental learning involves updating the pool of base classifiers by adding new members to this set. These new members are created with the Learn++ algorithm. In contrast, global incremental learning consists of updating the set of output profiles used during generalization. The proposed framework has been evaluated on a diversified set of databases. The results indicate that LoGID is promising. In most databases, the recognition rates achieved by the proposed method are higher than those achieved by other state-of-the-art approaches, such as batch learning. Furthermore, the simulated incremental learning setting demonstrates that LoGID can effectively improve the performance of systems created with small training sets as more data are observed over time.