A New Self-Organizing Map with Continuous Learning Capability
Hiroomi Hikawa, Hidetaka Ito, Yutaka Maeda · 2018
This paper proposes a new neighborhood function for the self-organizing map (SOM). As the learning of the SOM progresses, the conventional neighborhood function reduces its magnitude and neighborhood radius, and the learning stops after pre-defined training iterations. On the other hand, the proposed neighborhood function uses only the distance between the weight vector of the winner neuron and the input vector, then the magnitude and radius are computed according to this distance. Since the proposed neighborhood function is not a function of the learning iterations, it allows the SOM to continue its learning without stopping, and it can handle varying input vector distribution. This feature is especially effective under the unknown, dynamically changing input vector space that arises in, e.g., online learning. The proposed function uses vector distance to provide the SOM with voluntary learning capability. Therefore the vector distance plays a role of curiosity of biological brain.