A practical issue of edge AI systems is that data distributions of trained dataset and deployed environment may differ due to noise and environmental changes over time. Such a phenomenon is known as a concept drift, and this gap degrades the performance of edge AI systems and may introduce system failures. To address this gap, retraining of neural network models triggered by concept drift detection is a practical approach. However, since available compute resources are strictly limited in edge devices, in this paper we propose a fully sequential concept drift detection method in cooperation with an on-device sequential learning technique of neural networks. In this case, both the neural network retraining and the proposed concept drift detection are done only by sequential computation to reduce computation cost and memory utilization. We use three datasets for experiments and compare the proposed approach with existing batch-based detection methods. It is also compared with a DNN-based approach without concept drift detection. The evaluation results of the proposed approach show that the proposed method is capable of detecting each of four concept drift types. The results also show that, while the accuracy is decreased by up to 0.9% compared to the existing batch-based detection methods, it decreases the memory size by 88.9%-96.4% and the execution time by 45.0%-87.6%. As a result, the combination of the neural network retraining and the proposed concept drift detection method is demonstrated on Raspberry Pi Pico that has 264 kB memory.

In data stream analysis, detecting the concept drift accurately is important to maintain the classification performance. Most drift detection methods assume that the class labels become available immediately after a data sample arrives. However, it is unrealistic to attempt to acquire all of the labels when processing the data streams, as labeling costs are high and much time is needed. In this paper, we propose a concept drift detection method under the assumption that there is limited access or no access to class labels. The proposed method detects concept drift on unlabeled data streams based on the class label information which is predicted by a classifier or a virtual classifier. Experimental results on synthetic and real streaming data show that the proposed method is competent to detect the concept drift on unlabeled data stream.

In processing stream data, time is one of the most significant facts not only because the size of data is dramatically increased but because the characteristics of data is varying over time. To learn stream data evolving over time effectively, it is required to detect the drift of concept. We present a window adaptation function on domain value (WAV) to determine the size of windowed batch for learning algorithms of stream data and a method to detect the change of data characteristics with a criterion function utilizing correlation. When applying our adaptation function to a clustering task on a multi-stream data model, the result of learning synopsis of windowed batch determined by it shows its effectiveness. Our criterion function with correlation information of value distribution over time can be the reasonable threshold to detect the change between windowed batches.