Parallelized Convolutions for Embedded Ultra Low Power Deep Learning SoC

Deep Convolutional Neural Networks (DCNNs) achieve state of the art results compared to classic machine learning in many applications that need recognition, identification and classification. An ever-increasing embedded deployment of DCNNs inference engines thus supporting the intelligence close to the sensor paradigm has been observed, overcoming limitations of cloud-based computing as bandwidth requirements, security, privacy, scalability, and responsiveness. However, increasing the robustness and accuracy of DCNNs comes at the price of increased computational cost. As result, implementing CNNs on embedded devices with real-time constraints is a challenge if the lowest power consumption shall be achieved. A solution to the challenge is to take advantage of the intra-device massive fine grain parallelism offered by these systems and benefit from the extensive concurrency exhibited by DCNN processing pipelines. The trick is to divide intensive tasks into smaller, weakly interacting batches subject to parallel processing. Referred to that, this paper has mainly two goals: 1) describe the implementation of a state-of-art technique to map DCNN most intensive tasks (dominated by multiply-and-accumulate ops) onto Orlando SoC, an ultra-low power heterogeneous multi cores developed by STMicroelectronics; 2) integrate the proposed implementation on a toolchain that allows deep learning developers to deploy DCNNs on low-power applications