Development of Image Classification System Deploying Memory-Centric AI with High-capacity Storage

In deep learning, neural networks can acquire various knowledge and abilities through a training process, in which the parameters of the neural networks are tuned. On the other hand, if a user obtains new data (e.g., new images) after the training is completed and wants the neural network to acquire additional knowledge about that data, re-training is required. Re-training from scratch is costly in terms of time and power consumption. Therefore, a method of re-training with only a part of the data including new data, could be considered, but catastrophic forgetting problem, in which the knowledge and abilities learned in the past are lost, is inevitable.

We have addressed catastrophic forgetting problem in image classification tasks with an AI technology utilizing high-capacity storage (Memory-Centric AI). We devised a system that stores large amounts of image data, labels, and image feature maps^*1 as knowledge in the high-capacity storage, and that refers to the storage during image classification (Fig.1). In our system, when a user obtains new images, knowledge can be added or updated by storing additional image feature maps and labels without re-training or retuning parameters (Fig.2). Since the parameters of the neural network are not retuned at all, the catastrophic forgetting problem can be avoided. The knowledge can be added or updated while achieving better accuracy than those of conventional methods (Fig.3).

In addition, our system is expected to improve the explainability of AI^*2 and alleviate the black box problem^*3. In the conventional methods, when images are classified by neural networks, the process and basis for the results cannot be interpreted by humans. On the other hand, our system can visualize the basis for the classification results by listing the data referred to in the storage when classifying images (Fig.4). Furthermore, by analyzing the referred data, the contribution of each stored data can be evaluated according to the frequency of references.