Training Neural Networks for Sequential Change-point Detection

Detecting an abrupt distributional shift of the data stream, known as change-point detection, is a fundamental problem in statistics and signal processing. We present a new approach for online change-point detection by training neural networks (NN), and sequentially cumulating the detection statistics by evaluating the trained discriminating function on test samples by a CUSUM recursion. The idea is based on the observation that training neural networks through logistic loss may lead to the log-likelihood function. We demonstrated the good performance of NN-CUSUM on detecting change-point in high-dimensional data using both synthetic and real-world data.