A4 Vertaisarvioitu artikkeli konferenssijulkaisussa
TNEST: Training Sparse Neural Network for FPGA Based Edge Application
Tekijät: Das, Rammi; Karn, Rupesh Raj; Heikkonen, Jukka; Kanth, Rajeev
Toimittaja: Daimi Kevin, Al Sadoon Abeer
Konferenssin vakiintunut nimi: International Conference on Advances in Computing Research
Kustantaja: Springer Science and Business Media Deutschland GmbH
Kustannuspaikka: Cham
Julkaisuvuosi: 2024
Journal: Lecture notes in networks and systems
Kokoomateoksen nimi: Proceedings of the Second International Conference on Advances in Computing Research (ACR’24)
Tietokannassa oleva lehden nimi: Lecture Notes in Networks and Systems
Sarjan nimi: Lecture Notes in Networks and Systems
Vuosikerta: 956
Aloitussivu: 15
Lopetussivu: 28
ISBN: 978-3-031-56949-4
eISBN: 978-3-031-56950-0
ISSN: 2367-3370
eISSN: 2367-3389
DOI: https://doi.org/10.1007/978-3-031-56950-0_2
Verkko-osoite: https://doi.org/10.1007/978-3-031-56950-0_2
Machine learning (ML) hardware inference has developed ultra-low-power edge devices that accelerate inferential applications performance. An FPGA (Field Programmable Gate Array) is a popular option for such systems. The FPGA has power budget, memory, compute resource, area, etc. constraints but possesses several key advantages, including bandwidth saving, speed, real-time inference, offline activity, etc. Neural networks have been extensively used in Edge systems due to their significant popularity in AI. Requirements of complex neural networks have only been recently realized for edge applications, so the research community has yet to develop a standard model for such applications. The sparse neural architecture reduces the active neurons and connections, making the entire system computationally more efficient and utilizing less memory. Such essential facts are more and more evident in edge-based IoT applications. In this work, we have customized neural network training algorithms to fit precisely for the Edge systems. Rather than a traditional top-down approach where MLs are fully trained and then pruned to fit within edge device resources, we adopted a generative approach where MLs are prepared with the least number of parameters, and further components are added as the need arises to improve inference accuracy. Our generative model shows significant savings in FPGA resource consumption compared to the top-down approach for the same precision.
