Результаты исследований: Публикации в книгах, отчётах, сборниках, трудах конференций › статья в сборнике материалов конференции › научная › Рецензирование
Trainable Thresholds for Neural Network Quantization. / Goncharenko, Alexander; Denisov, Andrey; Alyamkin, Sergey и др.
Advances in Computational Intelligence - 15th International Work-Conference on Artificial Neural Networks, IWANN 2019, Proceedings. ред. / Ignacio Rojas; Gonzalo Joya; Andreu Catala. Springer, 2019. стр. 302-312 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Том 11507 LNCS).Результаты исследований: Публикации в книгах, отчётах, сборниках, трудах конференций › статья в сборнике материалов конференции › научная › Рецензирование
}
TY - GEN
T1 - Trainable Thresholds for Neural Network Quantization
AU - Goncharenko, Alexander
AU - Denisov, Andrey
AU - Alyamkin, Sergey
AU - Terentev, Evgeny
PY - 2019/1/1
Y1 - 2019/1/1
N2 - Embedded computer vision applications for robotics, security cameras, and mobile phone apps require the usage of mobile neural network architectures like MobileNet-v2 or MNAS-Net in order to reduce RAM consumption and accelerate processing. An additional option for further resource consumption reduction is 8-bit neural network quantization. Unfortunately, the known methods for neural network quantization lead to significant accuracy reduction (more than 1.2%) for mobile architectures and require long training with quantization procedure. To overcome this limitation, we propose a method that allows to quantize mobile neural network without significant accuracy loss. Our approach is based on trainable quantization thresholds for each neural network filter, that allows to accelerate training with quantization procedure up to 10 times in comparison with the standard techniques. Using the proposed technique, we quantize the modern mobile architectures of neural networks with the accuracy loss not exceeding 0.1%. Ready-for-use models and code are available at: https://github.com/agoncharenko1992/FAT-fast-adjustable-threshold.
AB - Embedded computer vision applications for robotics, security cameras, and mobile phone apps require the usage of mobile neural network architectures like MobileNet-v2 or MNAS-Net in order to reduce RAM consumption and accelerate processing. An additional option for further resource consumption reduction is 8-bit neural network quantization. Unfortunately, the known methods for neural network quantization lead to significant accuracy reduction (more than 1.2%) for mobile architectures and require long training with quantization procedure. To overcome this limitation, we propose a method that allows to quantize mobile neural network without significant accuracy loss. Our approach is based on trainable quantization thresholds for each neural network filter, that allows to accelerate training with quantization procedure up to 10 times in comparison with the standard techniques. Using the proposed technique, we quantize the modern mobile architectures of neural networks with the accuracy loss not exceeding 0.1%. Ready-for-use models and code are available at: https://github.com/agoncharenko1992/FAT-fast-adjustable-threshold.
KW - Distillation
KW - Machine learning
KW - Neural networks
KW - Quantization
UR - http://www.scopus.com/inward/record.url?scp=85067573043&partnerID=8YFLogxK
U2 - 10.1007/978-3-030-20518-8_26
DO - 10.1007/978-3-030-20518-8_26
M3 - Conference contribution
AN - SCOPUS:85067573043
SN - 9783030205171
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 302
EP - 312
BT - Advances in Computational Intelligence - 15th International Work-Conference on Artificial Neural Networks, IWANN 2019, Proceedings
A2 - Rojas, Ignacio
A2 - Joya, Gonzalo
A2 - Catala, Andreu
PB - Springer
T2 - 15th International Work-Conference on Artificial Neural Networks, IWANN 2019
Y2 - 12 June 2019 through 14 June 2019
ER -
ID: 20643808