Detection and Quantization of Data Drift in Image Classification Neural Networks

Danushka Senarathna, Spyros Tragoudas, Kiriti Nagesh Gowda, Mike Schmit

Abstract: An unforeseen change in the input data is called drift and may impact the accuracy of machine-learning models. A novel scheme for diagnosing data drift in the input stream of image classification neural networks is presented. The proposed drift detection and quantization method uses a threshold dictionary for the prediction probabilities of each class in the neural network model. The method is applicable to any drift type in images such as noise, and weather effects, among others. Experimental results on various datasets, drift types, and neural network models show that the proposed method estimates the drift magnitude with high accuracy, especially when the level of drift impacts the model's performance significantly.