Application of Learning to Learn to Real-World Pattern Recognition

Steve G. Romaniuk

Abstract: Pattern recognition has for decades played an important role in the development of intelligent systems and numerous algorithms have been proposed to deal with this important task. Though much interest and research has been invested into the construction of advanced pattern recognition systems, almost all of these algorithms are incapable of meta-learning. These one-shot learning systems do not address the problem of automatic adaptation of learning bias and hence lack one of the most fundamental requirements inherent to any truly autonomous pattern recognition system: the ability to learn to learn. The purpose of this paper is to report the findings of an extensive empirical study (i.e., including 26 real-world data sets) involving a system capable of meta-learning. Statistically significant reductions in both network units and training time are observed when utilizing a multi-shot learning environment.