Name of the participant: Arya Mazaheri
Description of the IT research project: Deep convolutional neural networks (ConvNet) is a well-known technique that has been successfully applied to many computer-vision tasks, such as object detection, semantic segmentation, and video classification. However, cost, power consumption, portability, and mass footprint are among product manufacturers’ main concerns during the training and deployment of such networks, requiring extensive optimizations to become suitable for industrial applications.
Training ConvNets are inherently highly parallelizable and heavily use computational resources, ranging from a high-end GPU to a large cluster of machines. Although such resources are readily available to the industry and product developers, not enough attention is being paid to the orchestration of AI jobs. Hence, AI projects have become the primary source of bottlenecks in research and development. As a remedy, we propose an AI-tailored job scheduler that can immensely enhance the scalability of the deep-learning training jobs, reduce the time-to-market, and save a considerable amount of operating costs.
oreover, once the training of a ConvNet converges to an optimum level, the obtained model is expected to be deployed on low-budget processors often found on edge devices. Nonetheless, deploying ConvNet models on such devices and exploiting the utmost performance is a cumbersome task, often demanding iterative rounds of tuning and code specialization. This project aims to focus on the computer system’s side of AI deployments and create efficient deep-learning vision algorithms for low-budget processors. First, we find an appropriate baseline ConvNet architecture suitable for scaling according to the target hardware architecture. Then, we generate hardware-aware ConvNet models using neural architecture search (NAS) by attempting to strike a balance between accuracy and runtime performance.
One of the missions of the German Federal Government within High-Tech Strategy 2025 is to put AI into practical application. The proposed project is on par with the same vision and aims to enable innovations in AI by exploring the path of deploying AI on millions, if not billions, of AI devices across the world. Particularly, this project enables affordable computer vision for all sorts of mass-produced devices, such as autonomous vehicles (cars, trains, buses, trucks), UAVs (Unmanned Aerial Vehicle, a.k.a. drones), and IoT (Internet of Things) devices.
Software Campus partners: TU Darmstadt, Huawei
Implementation period: 01.04.2021 – 31.03.2023
