Binary Neural Networks (BNN) have been proposed to address the computational complexity and memory requirements of Convolutional Neural Networks (CNN). However, in most of the applications, BNNs suffer from severe accuracy loss due to the 1-bit quantization. In this paper, a Mixed-Precision Binary Weight Network (MP-BWN) is proposed as a compromise between CNN and BNN. Compared to traditional binary networks, MP-BWN offers better performance with an acceptable increase in the network size. MP-BWN achieves up to 99% reduction in both the number of operations and the network size compared to similar state-of-the-art solutions. When validated on a touch modality classification problem, the MP-BWN surpassed similar existing solutions by achieving a classification accuracy of 77.8%.

A mixed-precision binary neural network architecture for touch modality classification

Younes H.;Ibrahim A.;Valle M.
2021-01-01

Abstract

Binary Neural Networks (BNN) have been proposed to address the computational complexity and memory requirements of Convolutional Neural Networks (CNN). However, in most of the applications, BNNs suffer from severe accuracy loss due to the 1-bit quantization. In this paper, a Mixed-Precision Binary Weight Network (MP-BWN) is proposed as a compromise between CNN and BNN. Compared to traditional binary networks, MP-BWN offers better performance with an acceptable increase in the network size. MP-BWN achieves up to 99% reduction in both the number of operations and the network size compared to similar state-of-the-art solutions. When validated on a touch modality classification problem, the MP-BWN surpassed similar existing solutions by achieving a classification accuracy of 77.8%.
File in questo prodotto:
File Dimensione Formato  
PRIME2021_submitted.pdf

accesso chiuso

Descrizione: Contributo in atti di convegno
Tipologia: Documento in Post-print
Dimensione 281.91 kB
Formato Adobe PDF
281.91 kB Adobe PDF   Visualizza/Apri   Richiedi una copia

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11567/1069128
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 0
  • ???jsp.display-item.citation.isi??? ND
social impact