An importance weighted projection method for incremental learning under unstationary environments

In this paper, we propose a new projection method for incremental learning on a fixed number of kernels. If the number of the kernels reaches the upper bound, the learning machine has to dispose of part of the memory in varying degrees to make space for the recording of a new instance. If we assume that the environment is ergodic, where the learned samples will appear again later, the learning machine should minimize the disposing ratio to yield a correct response when it encounters the learned samples. To achieve this goal, we reconstruct a kernel-based projection method that minimizes the magnitude of forgetting as well as the current error to the new instance. Next, the method is extended to take into account the sample distribution. The experimental results show that the proposed method is superior to other kernel based learning methods for minimizing the mean square error of the given samples.