In conventional AI techniques, neural networks are trained to acquire knowledge by updating parameters called “weights.” Once fully trained, in order to acquire new knowledge a neural network must be either re-trained from the beginning or fine-tuned with new data. The former requires huge amounts of time and consumes significant energy costs, while the latter requires parameters to be updated and faces the catastrophic forgetting problem of losing the knowledge acquired in the past which leads to deterioration of classification accuracy.

To address the issues of cost and accuracy in neural network-based image classification systems, the new solution stores large amounts of image data, labels and image feature maps as knowledge in a high-capacity storage. The neural network then classifies images by referring to that stored knowledge (Figure). Using this method, knowledge can be added or updated by adding newly obtained image labels and feature maps to the stored data. As there is no need to re-train or update weights, which may cause “catastrophic forgetting,” image classification can be maintained more accurately.

Furthermore, by using the data referred from the storage when the neural network classifies images, the basis for the classification results can be visualized, which is expected to improve the explainability of AI[3] and alleviate the black-box problem[4], in turn allowing the selective modification of knowledge sources. In addition, by analyzing the referred data, the contribution of each stored data can be evaluated according to the frequency of references.

Guided by its mission of "Uplifting the World with ‘Memory,’" Kioxia will continue to contribute to the development of AI and storage technologies by expanding Memory-Centric AI beyond image classification to other areas and promoting research and development of AI technology utilizing high-capacity storage.