MACHINE PERCEPTION AND ROBOTICS GROUP

機械知覚&ロボティクスグループ
中部大学

Data drift is a change in the feature distribution of input data during machine learning model training and during system operation. The data drift occurs regardless of the type of data and adversely affects model performance. The existing methods detects the data drift by using two-sample test for network output. However, these methods merely apply two-sample test with the distribution of class probabilities. Even though drifted input images are transformed by noise and/or geometric transformations, these methods does not consider such transformations. In addition to class probability, we believe that detecting drift for changes in the local region that the model is actually gazing at will improve accuracy. In this study, we propose a drift detection method based on attention branch network (ABN), which enables visualization of the basis of judgment in image classification. In our method, drift is detected using the class probabilities output by the attention branch and perception branch, which constitute the ABN, and the attention map. The results show that we can improve the detection ratio by introducing an attention map to drift detection in addition to class probability. We also observed that the attention map tended to shrink with drift.