Connectionist systems for rapid adaptive learning: A comparative analysis on speech recognition

In may real time applications a system needs to be trained quickly on a small amount of new data and then generalise on unseen data. This is a challenging task for neural networks especially when applied to difficult problems such as adaptive speech recognition. An experimental comparison between three recently introduced on-line adaptive connectionist paradigms on adaptive phoneme recognition task is presented in the paper. The models used are: the Evolving Classifying Function (ECF), the Zero Instruction Set Computer (ZISC) and the Deterministic Adaptive Random Access Memory Network (DARN). An MLP network, although slow and not an incremental learning model, was used to provide a benchmark for comparison. The results show that ECF adapts faster to new data, while ZISC and DARN require mode data.