Artigo em revista científica Q2
ULEEN: A novel architecture for ultra low-energy edge neural networks
Zachary Susskind (Susskind, Z.); Aman Arora (Arora, A.); Igor D. S. Miranda (Miranda, I. D. S.); Alan T. L. Bacellar (Bacellar, A. T. L.); Luis A. Q. Villon (Villon, L. A. Q.); Rafael F. Katopodis (Katopodis, R. F.); Leandro Santiago de Araújo (Araújo, L. S. de); Diego Leonel Cadette Dutra (Dutra, D. L. C.); Priscila M.V. Lima (Lima, P. M. V. L.); Felipe Maia Galvão França (França, F.); Maurício Breternitz (Breternitz, M.); Lizy K. John (John, L. K.); et al.
Título Revista
ACM Transactions on Architecture and Code Optimization
Ano (publicação definitiva)
2023
Língua
Inglês
País
Estados Unidos da América
Mais Informação
Web of Science®

N.º de citações: 1

(Última verificação: 2024-11-04 00:50)

Ver o registo na Web of Science®


: 0.7
Scopus

N.º de citações: 4

(Última verificação: 2024-11-01 14:12)

Ver o registo na Scopus


: 2.0
Google Scholar

N.º de citações: 8

(Última verificação: 2024-11-02 13:33)

Ver o registo no Google Scholar

Abstract/Resumo
"Extreme edge"1 devices, such as smart sensors, are a uniquely challenging environment for the deployment of machine learning. The tiny energy budgets of these devices lie beyond what is feasible for conventional deep neural networks, particularly in high-throughput scenarios, requiring us to rethink how we approach edge inference. In this work, we propose ULEEN, a model and FPGA-based accelerator architecture based on weightless neural networks (WNNs). WNNs eliminate energy-intensive arithmetic operations, instead using table lookups to perform computation, which makes them theoretically well-suited for edge inference. However, WNNs have historically suffered from poor accuracy and excessive memory usage. ULEEN incorporates algorithmic improvements and a novel training strategy inspired by binary neural networks (BNNs) to make significant strides in addressing these issues. We compare ULEEN against BNNs in software and hardware using the four MLPerf Tiny datasets and MNIST. Our FPGA implementations of ULEEN accomplish classification at 4.0-14.3 million inferences per second, improving area-normalized throughput by an average of 3.6× and steady-state energy efficiency by an average of 7.1× compared to the FPGA-based Xilinx FINN BNN inference platform. While ULEEN is not a universally applicable machine learning model, we demonstrate that it can be an excellent choice for certain applications in energy- and latency-critical edge environments.
Agradecimentos/Acknowledgements
--
Palavras-chave
Weightless neural networks,WiSARD,Neural networks,Inference,Edge computing,MLPerf tiny,High throughput computing
  • Ciências da Computação e da Informação - Ciências Naturais
Registos de financiamentos
Referência de financiamento Entidade Financiadora
DSAIPA/AI/0122/2020 Fundação para a Ciência e a Tecnologia
POCI-01-0247-FEDER-045912 Comissão Europeia
3148.001 Semiconductor Research Corporation
3015.001 Semiconductor Research Corporation
2326894 National Science Foundation
UIDP/04466/2020 Fundação para a Ciência e a Tecnologia
UIDB/50008/2020 Fundação para a Ciência e a Tecnologia
UIDB/04466/2020 Fundação para a Ciência e a Tecnologia