Towards Efficient In-memory Computing Hardware for Quantized Neural Networks: State-of-the-art, Open Challenges and Perspectives

The amount of data processed in the cloud, development Internet-of-Things (IoT) applications, and growing privacy concerns force transition from cloud-based to edge-based processing. Limited energy computational resources on edge push traditional von Neumann architectures In-memory Computing (IMC), especially for machine learning neural network applications. Network compression techniques are applied implement a limited hardware resources. Quantization is one most efficient allowing reduce memory footprint, latency, consumption. This paper provides comprehensive review IMC-based Quantized Neural Networks (QNN) links software-based quantization approaches IMC implementation. Moreover, open challenges, QNN design requirements, recommendations, perspectives along with an roadmap provided.