A dictionary updating scheme incorporating words elimination into Quantized Kernel Least-Mean-Squares for changing environments

Learning under time-varying environment is a challenging task since one has to deal with the ever changing distribution of data. A common and yet effective solution is to learn the data online and keep up with any ongoing changes. The Quantized Kernel Least-Squares (QKLMS) is an effective tool for online dictionary learning where the network size is capped by the quantization dictionary size. However, due to the lack of a mechanism to eliminate outdated words, learning can become irrelevant over time. In this paper, a mechanism to remove irrelevant words in the dictionary is proposed for QKLMS. Our experimental results based on chaotic time sequence prediction validate the capability of the developed method for time-varying data adaptation.