Duration: 24 months
Deep Neural Networks (DNNs) are becoming viable for diverse space applications such as Earth observation [1] and spacecraft pose estimation [2]. Deployment in space, however, raises the question of resilience to memory and input errors. Common steps toward resilience include fault-tolerant training [3-6] and ensemble models [7,8]. These techniques, however, do not consider the impact of the DNN architecture on fault tolerance. This is the overarching goal of this project. Considering the memory and computation constraints due to space deployment, we will investigate the influence of DNN compression strategies on error resilience. While a few works have studied fault tolerance for pruned or quantized DNNs, they have reached contradictory conclusions, showing either improved [9,10] or degraded [11,12] error resilience. We argue that these inconsistencies are due to the limited scope of these studies. In other words, they evidence that the impact of DNN compression is intertwined with the DNN architecture and the task at hand. In this project, we will investigate via 3 main axes: (i) A large-scale analysis of the effect of different quantization and knowledge distillation strategies, including our own [13-15], on fault tolerance, using a variety of DNN architectures and datasets; (ii) the incorporation of diverse faults occurring within the network and at the input, culminating in a space-environment fault injection pipeline; (iii) the study of the impact of hardware on fault tolerance, incorporating it in the compression strategy. In contrast to most studies focussed on image recognition [8-12,16-18], we will address spacecraft pose estimation, which has clear industry applications for on-orbit servicing and debris removal (e.g., Clearspace, Astroscale, Thales Alenia Space etc.). Thanks to access to ESA hardware platforms and testing facilities, validation will be carried out on diverse hardware (CPU, FPGA, GPU) using Klepsydra‘s software inference tools.
