Optimizing DNN Architectures for High Speed Autonomous Navigation in GPS Denied Environments on Edge Devices

Abstract

We address the challenge of high speed autonomous navigation of micro aerial vehicles (MAVs) using DNNs in GPS-denied environments with limited computational resources; specifically, we use the ODROID XU4 and the Raspberry Pi 3. The high computation costs of using DNNs for inference, particularly in the absence of powerful GPUs, necessitates negotiating a tradeoff between accuracy and inference. We address this tradeoff by employing sparsified neural networks. To obtain such architectures, we propose a novel algorithm to find sparse “sub networks” of existing pre trained models. Contrary to existing pruning-only strategies, our proposal includes a novel exploration step that efficiently searches for a different, but identically sparse, architecture with better generalization abilities. We derive learning theoretic bounds that reinforce our empirical findings that the optimized network achieves comparable generalization to the original network. We show that using our algorithm it is possible to discover models which, on average, have upto 19x fewer parameters than those obtained using existing state of the art pruning methods on autonomous navigation datasets, and achieve upto 6x improvements on inference time compared to existing state of the art shallow models on the ODROID XU4 and Raspberry Pi 3. Last, we demonstrate that our sparsified models can complete autonomous navigation missions with speeds upto 4 m/s using the ODROID XU4, which existing state of the art methods fail to do.