uPIM: Performance-aware Online Learning Capable Processing-in-Memory

Machine learning and AI-based automated systems are gaining increasing attention for real-time intelligent applications by virtue of a superior co-ordination between the software and the hardware within these systems. Although the majority of the automated systems are implementing Convolutional neural networks (CNNs), and Deep Neural Networks (DNNs) on the hardware with impressive accuracy, a significant amount of cost is associated with data movement in these platforms. Recent advancements in processing-in-memory (PIM), a non-von Neumann computing paradigm, have proven to be very effective in minimizing data communication overheads by performing computations within the memory chip. However, these devices are primarily designed as inference engines and therefore have not been adequately investigated for real-time learning capabilities for applications in changing environments. In this work, we introduce uPIM, a PIM architecture that supports a Generative Adversarial Network (GAN)-based performance-aware online learning model for updating the weights with minimal overheads. Our hardware-software co-design approach exhibits superior performance and efficiency in real-time applications like Autonomous Navigation Systems (ANS) by leveraging massive data-level parallelism and ultra-low data movement latency. The evaluations are performed on multiple state-of-the-art deep learning networks like LeNet, AlexNet, ResNet18, 34, 50 on the German Traffic Sign Recognition Benchmark (GTSRB) dataset and the Belgium Traffic Sign Dataset (BTSD) with several data-precisions. The proposed performance-aware, quantization-friendly online learning based PIM architecture achieves an average accuracy of 72% for GTSRB and 83.4% for BTSD dataset under varying environment for CNNs implemented for Traffic Sign Recognition (TSR) with 8-bit fixed point data-precision.