Conferência De: orcid, cienciavitae

PositNN: Training Deep Neural Networks with Mixed Low-Precision Posit

ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

Gonçalo Raposo; Pedro Tomás; Nuno Roma — 2021 — IEEE

Informações chave

Autores:

Gonçalo Raposo (Gonçalo Eduardo Cascalho Raposo); Pedro Tomás (Pedro Filipe Zeferino Tomás); Nuno Roma (Nuno Filipe Valentim Roma)

Publicado em

Junho 2021

Resumo

Low-precision formats have proven to be an efficient way to reduce not only the memory footprint but also the hardware resources and power consumption of deep learning computations. Under this premise, the posit numerical format appears to be a highly viable substitute for the IEEE floating-point, but its application to neural networks training still requires further research. Some preliminary results have shown that 8-bit (and even smaller) posits may be used for inference and 16-bit for training, while maintaining the model accuracy. The presented research aims to evaluate the feasibility to train deep convolutional neural networks using posits. For such purpose, a software framework was developed to use simulated posits and quires in end-to-end training and inference. This implementation allows using any bit size, configuration, and even mixed precision, suitable for different precision requirements in various stages. The obtained results suggest that 8-bit posits can substitute 32-bit floats during training with no negative impact on the resulting loss and accuracy.

Detalhes da publicação

Autores da comunidade :

Versão da publicação

SMUR - Versão submetida a revisão

Editora

IEEE

Ligação para a versão da editora

https://ieeexplore.ieee.org/document/9413919

Título do contentor da publicação

ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

Local da conferência

Canada

Data de início conferência

06/06/2021

Data de término da conferência

11/06/2021

ISBN

978-1-7281-7605-5

ISSN

1520-6149

DOI

10.1109/ICASSP39728.2021.9413919

Domínio Científico (FOS)

electrical-engineering-electronic-engineering-information-engineering - Engenharia Eletrotécnica, Eletrónica e Informática

Palavras-chave

  • Posit numerical format
  • low-precision arithmetic
  • deep neural networks
  • training
  • inference

Idioma da publicação (código ISO)

eng - Inglês

Acesso à publicação:

Acesso Aberto