Recursive total error rate minimization

In learning algorithm design, a direct optimization of a cost function which fits the goal is naturally desired. Particularly in recursive learning, a direct formulation for total-error-rate (TER) minimization is much desired for online classification applications. However, due to a nonlinear counting step in the classification formulation, an exact solution to minimize TER recursively is yet to be established. In this paper, we propose an exact recursive formulation for TER minimization. Based on empirical evaluations using benchmark data sets, we show that the proposed recursive classification algorithm preserves the performance of the batch mode TER while easing the computational memory load by sample based accumulation.