An initial parameter search for rapid concept drift adaptation in deep neural networks

Abstract

Concept drift is a common issue in data stream mining algorithms that causes prediction models to lose its original performance gradually or abruptly due to the non-stationarity of the data distribution and decision boundaries. To combat concept drifts, prediction models need to be updated periodically or when concept drifts occur to adapt to the current concept. Unfortunately, training deep neural networks often require a large amount of data samples and high computational resource consumption, making adaptation slow when concept drifts occur. This paper proposes an approach by searching for an optimum initial parameter that could be adapted quickly to all possible concept drift situations. The initial parameter search is based on the Reptile [1] algorithm, which had been successfully applied in image classification, which allows a neural network model to learn from a few samples and minimal gradient steps. We argue that using an optimum initial parameter allows prior information to be embedded and makes the prediction model less reliant on training exclusively from new data when concept drift occurs. Experimental results show that this approach performs at least as well as current data streaming algorithms but with the lowest computational overhead.