Extending dynamic SOMs to capture incremental changes in data

Humans learn in an incremental manner. Due to this reason, humans continuously refine their knowledge of the environment with the experience gained. Many strides have been made in the machine learning area to exploit the power of incremental learning. Incremental learning, in contrast to onetime learning is far more useful and effective when data is not completely available at once. Here, we investigate an unsupervised incremental learning algorithm known as Incremental Knowledge Acquisition and Self Learning (IKASL) algorithm. IKASL algorithm is able to capture knowledge in an incremental manner, without disrupting past knowledge. Furthermore, IKASL algorithm encodes acquired knowledge in such a way that it can be used to acquire new knowledge more efficiently. This paper discusses several limitations of original IKASL algorithm and proposes several extensions to the original algorithm which enhances its performance. These modifications include influencing spread factor, implementing fuzzy integral based generalizing technique, etc. Furthermore, the paper report observations of several experiments conducted with several datasets to assess the necessity and value of incremental learning in the real world. These experiments are carefully designed to reflect interesting characteristics of the IKASL algorithm.