Method for Adaptive Training of Polynomial Networks with Applications to Speaker Verification

Speaker verification is the process of determining the validity of a claimed identity through voice. Traditional approaches to this problem are Gaussian mixture models and hidden Markov models. Although these methods work well, they are difficult to employ in an adaptive framework because of the iterative nature of training. Ideally, as we acquire new-labeled input, we would like to update the verification model immediately to avoid storing speech data (for small memory situations) and to adapt to speaker variability. In this paper, we propose a novel method for adaptive training of polynomial networks. We show that the new method is computationally efficient, requires little memory, and is competitive with batch-based training.