Low Power Deep-Learning Architecture for Mobile IoT Intelligence

We present and develop DeepIoT, a low power convolutional neural networks (CNNs) accelerator to enhance inference intelligence on low-end mobile Internet of Things (IoT) devices. The develop of DeepIoT architecture is motivated with the applications need to design an efficient CNN architecture which is efficient both in terms of power consumption and network model size. The proposal includes two techniques: a model compression based on hybrid quantization scheme and a binary activation function based on support vector machine (SVM). To validate the proposed solution, a VGG model with binary weights in fully connected (FC) layers and ternary weights for the convolution (CONV) layers and SVM activations is implemented on embedded FPGA platform. The evaluation shows the latency and power consumption performance of the proposed architecture is competitive compared with state-of-the-art reduced precision CNN models, such as VGG-16.