Machine learning technologies inspired by neural networks are likely to be most useful when modeling difficult-to-code tasks, such as gender recognition based on facial images. This paper summarizes some recent experiments with one such technology that appears to learn the gender recognition task quite well and quite efficiently.
Raafat, Tolba, and Shaddad use learning vector quantization (LVQ) neural networks that partition a training set into clusters of similar inputs and then categorize an input as belonging to a particular category, if it most closely matches one of the clusters in that category. By varying parameters, multiple instances of these networks can be constructed and trained, each performing differently than the others. This paper combines such instances into committee machines, in two different ways, so that the final classification is given by the majority vote of the component networks. The batch way effectively constructs and trains the component networks in parallel, while the adaptive way adds and trains more component networks, one at time, until performance stops improving.
Based on the results reported, these committee machines, especially the adaptive ones, perform the gender recognition task as well as humans and better than virtually all the other technologies with which they were compared. Readers with an interest in machine learning will find this paper interesting and comprehensible, even if they have little knowledge of neural networks.