Translating Feedforward Neural Nets to SOM-like Maps

A major disadvantage of feedforward neural networks is still the difficulty to gain insight into their internal functionality. This is much less the case for, e.g., nets that are trained unsupervised, such as Kohonen’s self-organizing feature maps (SOMs). These offer a direct view into the stored knowledge, as their internal knowledge is stored in the same format as the input data that was used for training or is used for evaluation. This paper discusses a mathematical transformation of a feed-forward network into a SOMlike structure such that its internal knowledge can be visually interpreted. This is particularly applicable to networks trained in the general classification problem domain. Keywords— Neural networks, rule extraction, selforganizing maps, feature maps, character recognition.