Binary Quantization in Convolutional Neural Network

Authors

  • Nicolás Urbano Pintos Grupo Tecnología Aplicada al Medio Ambiente, Facultad Regional Haedo, Universidad Tecnológica Nacional / División Radar Laser, Instituto de Investigaciones Científicas y Técnicas para la Defensa - Argentina
  • Mario Lavorato Director
  • Héctor Lacomi Codirector

DOI:

https://doi.org/10.33414/ajea.1132.2022

Keywords:

Deep Learning, Convolutional Neural Network, Binary Quantization

Abstract

In this work, it is proposed to implement a convolution Binarized Neural Network (BNN) to classify objects from RGB images. BNNs reduce the amount of computational and memory resources, and allow them to be inferred in embedded systems such as FPGAs, achieving real-time responses. The model is based on the VGG16 network and is trained on the CIFAR10 dataset. The network is binary quantized with the training-aware quantization technique (QAT). An accuracy close to 88% was achieved with the CIFAR10 evacuation set.

Downloads

Download data is not yet available.

Downloads

Published

2022-10-03

How to Cite

Urbano Pintos, N., Lavorato, M., & Lacomi, H. (2022). Binary Quantization in Convolutional Neural Network. AJEA (Proceedings of UTN Academic Conferences and Events), (15). https://doi.org/10.33414/ajea.1132.2022

Conference Proceedings Volume

No. 15 (2022): VI Conference on Exchange and Dissemination of Research Results of Doctoral Students in Engineering

Section

Proceedings - Signal and Image Processing