Structure adaptive multilayer overlapped SOMs with supervision for handprinted digit classification

We present a hybrid learning algorithm, structure adaptation techniques, and multilayered and overlapped structure, for the standard self-organising maps (SOM) to obtain an extremely powerful labelled pattern classification system. The learning algorithm consists of the standard unsupervised SOM learning of synaptic weights as well as a supervised learning of weights. The supervision stage is used to guide the structure adaptation process, to fine tune the weights and to obtain a network with good generalisation performance by avoiding over-training. In fact classifiers based on self-organising/unsupervised neural networks commonly suffer from over-training. As higher layer SOMs overlap, the final classification is made by fusing the classifications of individual overlapped SOMs. We obtained the best results ever reported for any SOM-based numerals classification system.