Processing In-Memory Architecture with On-Chip Transfer Learning Function for Compensating Characteristic Variation

Along with more awareness of the need to endpoint data processing, the endpoint artificial intelligence (AI) technology with transfer learning function that enables not only inference but also additional learning at endpoint for adapting to the difference of surrounding environment will be of much importance in the future. In this work, Flash memory-based Processing In-Memory architecture (Flash-PIM) with transfer learning function and its learning method are proposed. By sharing the flash memory array for storing connection weight of synapse during inference and learning process, total area is reduced by 25.1 % and abnormality detection accuracy is 1.9 - 2.3 % better than conventional transfer learning due to eliminating the influence of variations in Flash memory cells. This approach is also effective for other non-volatile memory devices such as emerging memories.