The original paper is in English. Non-English content has been machine-translated and may contain typographical errors or mistranslations. ex. Some numerals are expressed as "XNUMX".
Copyrights notice
The original paper is in English. Non-English content has been machine-translated and may contain typographical errors or mistranslations. Copyrights notice
Redes Neurais Binárias (BNN) binarizaram valores de neurônios e conexões para que seus aceleradores possam ser realizados por hardware extremamente eficiente. No entanto, existe uma lacuna significativa de precisão entre BNNs e redes com maior largura de bits. BNNs convencionais binarizam mapas de características por limites estáticos globalmente unificados, o que faz com que a imagem bipolar produzida perca detalhes locais. Este artigo propõe uma função de ativação de múltiplas entradas para permitir o limiar adaptativo para binarizar mapas de recursos: (a) No nível do algoritmo, em vez de operar cada pixel de entrada independentemente, o limiar adaptativo altera dinamicamente o limiar de acordo com os pixels circundantes do pixel alvo. Ao otimizar pesos, o limiar adaptativo é equivalente a uma convolução acompanhada em profundidade entre a convolução normal e a binarização. Os pesos acompanhados nos filtros de profundidade são ternarizados e otimizados de ponta a ponta. (b) No nível de hardware, o limiar adaptativo é realizado através de uma função de ativação de múltiplas entradas, que é compatível com arquiteturas de aceleradores comuns. Foi desenvolvido um hardware de ativação compacto com apenas um acumulador extra. Ao equipar o método proposto em FPGA, é alcançada uma melhoria de precisão de 4.1% no BNN original com apenas 1.1% de recurso LUT extra. Em comparação com os métodos de última geração, a ideia proposta aumenta ainda mais a precisão da rede em 0.8% no conjunto de dados Cifar-10 e 0.4% no conjunto de dados ImageNet.
Peiqi ZHANG
The University of Tokyo
Shinya TAKAMAEDA-YAMAZAKI
The University of Tokyo
The copyright of the original papers published on this site belongs to IEICE. Unauthorized use of the original or translated papers is prohibited. See IEICE Provisions on Copyright for details.
Copiar
Peiqi ZHANG, Shinya TAKAMAEDA-YAMAZAKI, "MITA: Multi-Input Adaptive Activation Function for Accurate Binary Neural Network Hardware" in IEICE TRANSACTIONS on Information,
vol. E106-D, no. 12, pp. 2006-2014, December 2023, doi: 10.1587/transinf.2023PAP0007.
Abstract: Binary Neural Networks (BNN) have binarized neuron and connection values so that their accelerators can be realized by extremely efficient hardware. However, there is a significant accuracy gap between BNNs and networks with wider bit-width. Conventional BNNs binarize feature maps by static globally-unified thresholds, which makes the produced bipolar image lose local details. This paper proposes a multi-input activation function to enable adaptive thresholding for binarizing feature maps: (a) At the algorithm level, instead of operating each input pixel independently, adaptive thresholding dynamically changes the threshold according to surrounding pixels of the target pixel. When optimizing weights, adaptive thresholding is equivalent to an accompanied depth-wise convolution between normal convolution and binarization. Accompanied weights in the depth-wise filters are ternarized and optimized end-to-end. (b) At the hardware level, adaptive thresholding is realized through a multi-input activation function, which is compatible with common accelerator architectures. Compact activation hardware with only one extra accumulator is devised. By equipping the proposed method on FPGA, 4.1% accuracy improvement is achieved on the original BNN with only 1.1% extra LUT resource. Compared with State-of-the-art methods, the proposed idea further increases network accuracy by 0.8% on the Cifar-10 dataset and 0.4% on the ImageNet dataset.
URL: https://global.ieice.org/en_transactions/information/10.1587/transinf.2023PAP0007/_p
Copiar
@ARTICLE{e106-d_12_2006,
author={Peiqi ZHANG, Shinya TAKAMAEDA-YAMAZAKI, },
journal={IEICE TRANSACTIONS on Information},
title={MITA: Multi-Input Adaptive Activation Function for Accurate Binary Neural Network Hardware},
year={2023},
volume={E106-D},
number={12},
pages={2006-2014},
abstract={Binary Neural Networks (BNN) have binarized neuron and connection values so that their accelerators can be realized by extremely efficient hardware. However, there is a significant accuracy gap between BNNs and networks with wider bit-width. Conventional BNNs binarize feature maps by static globally-unified thresholds, which makes the produced bipolar image lose local details. This paper proposes a multi-input activation function to enable adaptive thresholding for binarizing feature maps: (a) At the algorithm level, instead of operating each input pixel independently, adaptive thresholding dynamically changes the threshold according to surrounding pixels of the target pixel. When optimizing weights, adaptive thresholding is equivalent to an accompanied depth-wise convolution between normal convolution and binarization. Accompanied weights in the depth-wise filters are ternarized and optimized end-to-end. (b) At the hardware level, adaptive thresholding is realized through a multi-input activation function, which is compatible with common accelerator architectures. Compact activation hardware with only one extra accumulator is devised. By equipping the proposed method on FPGA, 4.1% accuracy improvement is achieved on the original BNN with only 1.1% extra LUT resource. Compared with State-of-the-art methods, the proposed idea further increases network accuracy by 0.8% on the Cifar-10 dataset and 0.4% on the ImageNet dataset.},
keywords={},
doi={10.1587/transinf.2023PAP0007},
ISSN={1745-1361},
month={December},}
Copiar
TY - JOUR
TI - MITA: Multi-Input Adaptive Activation Function for Accurate Binary Neural Network Hardware
T2 - IEICE TRANSACTIONS on Information
SP - 2006
EP - 2014
AU - Peiqi ZHANG
AU - Shinya TAKAMAEDA-YAMAZAKI
PY - 2023
DO - 10.1587/transinf.2023PAP0007
JO - IEICE TRANSACTIONS on Information
SN - 1745-1361
VL - E106-D
IS - 12
JA - IEICE TRANSACTIONS on Information
Y1 - December 2023
AB - Binary Neural Networks (BNN) have binarized neuron and connection values so that their accelerators can be realized by extremely efficient hardware. However, there is a significant accuracy gap between BNNs and networks with wider bit-width. Conventional BNNs binarize feature maps by static globally-unified thresholds, which makes the produced bipolar image lose local details. This paper proposes a multi-input activation function to enable adaptive thresholding for binarizing feature maps: (a) At the algorithm level, instead of operating each input pixel independently, adaptive thresholding dynamically changes the threshold according to surrounding pixels of the target pixel. When optimizing weights, adaptive thresholding is equivalent to an accompanied depth-wise convolution between normal convolution and binarization. Accompanied weights in the depth-wise filters are ternarized and optimized end-to-end. (b) At the hardware level, adaptive thresholding is realized through a multi-input activation function, which is compatible with common accelerator architectures. Compact activation hardware with only one extra accumulator is devised. By equipping the proposed method on FPGA, 4.1% accuracy improvement is achieved on the original BNN with only 1.1% extra LUT resource. Compared with State-of-the-art methods, the proposed idea further increases network accuracy by 0.8% on the Cifar-10 dataset and 0.4% on the ImageNet dataset.
ER -