Random sketch learning for deep neural networks in edge computing

Despite the great potential of deep neural networks (DNNs), they require massive weights and huge computational resources, creating a vast gap when deploying artificial intelligence at low-cost edge devices. Current lightweight DNNs, achieved by high-dimensional space pre-training post-compression, present challenges covering resources deficit, making tiny hard to be implemented. Here we report an architecture named random sketch learning, or Rosler, for computationally efficient intelligence. We build universal compressing-while-training framework that directly learns compact model and, most importantly, enables on-device learning. As validated on different models datasets, it attains substantial memory reduction ~50–90× (16-bits quantization), compared with fully connected DNNs. demonstrate hardware, whereby computation is accelerated >180× energy consumption reduced ~10×. Our method paves way in many scientific industrial applications. Developing networks, while essential computing, still remains challenge. Random learning creates thus paving machine (TinyML) resource-constrained