Indexat a

Llicència i ús

Citacions

Altmetrics

Anàlisi d'autories institucional

Compartir

Analysis and methodology for enabling DNN inference in an IoT edge environment in depth completion tasks

Publicat a:Dcis 2022 - Proceedings Of The 37th Conference On Design Of Circuits And Integrated Systems. 6-11 - 2022-01-01 (), DOI: 10.1109/DCIS55711.2022.9970054

Autors: Martinez de Ternero, Alejandro; Sancho, Jaime; Vazquez, Guillermo; Villa, Manuel; Rosa, Gonzalo; Sutradhar, Pallab; Martin-Perez, Alberto; Chavarrias, Miguel; Jimenez-Roldan, Luis; Perez-Nunez, Angel; Lagares, Alfonso; Juarez, Eduardo; Sanz, Cesar

Afiliacions

Resum

IoT and artificial intelligence (AI) have attracted a huge amount of interest in recent years, creating the so-called edge artificial intelligence field. However, due to the tendency in AI to create larger and more complex models, these algorithms are often not suitable for resource-restricted devices, and efforts have to be made in the direction of optimizing AI for IoT. This paper presents an analysis of three different tools that allow the use of Deep Neural Network (DNN) models in resource-limited devices: (i) a deep neural network compressor, (ii) the use of different bit-precision operations in the model, and (iii) the use of a specific accelerator engine in the constrained device. For that purpose, an IoT architecture is followed in which the edge device is a constrained device which takes the model pre-trained from the cloud and performs real-time inference. In particular, four common CNN building blocks have been compared in an encoder-decoder architecture, in order to improve the depth maps generated by LiDAR-like cameras in situ. The resulting architectures have been analyzed in terms of quality, energy consumption and inference time, measuring the effects of the compressor, the bit precision and the accelerator engine. The results obtained show that the use of a methodology based on these three tools combined is beneficial to reduce the size of the model more than ×8 times, energy consumption by almost ×3 times, and an increase in speed by ×3 times from the base model, without compromising the quality of the model.

Paraules clau