이진화된 컨벌루션 신경망 기반의 추론을 위한 효율적인 워크로드 분할 = Efficient Workload Division for Binarized-Convolutional-Neural-Network-Based Inference

다국어 초록 (Multilingual Abstract)

This paper studies the previous methods of the workload division for the inference based on the binary convolutional neural networks and proposes a novel method based on the output-channel-wise workload division. The proposed method, which divides the weight in the output-channel direction, reduces the off-chip memory access and does not entail additional computations. The experimental results for the ImageNet classification task show that the proposed method reduces the on-chip memory size and off-chip memory access by 89% and 88%, respectively, when compared to the previous method. Furthermore, no additional computations are entailed.

국문 초록 (Abstract)

본 논문은 이진화된 컨벌루션 신경망 기반의 추론을 위한 워크로드 분할 방식을 탐구하고, 출력 채널 방향으로의 분할 방식을 제안한다. 제안하는 워크로드 분할 방식은 공간 상의 출력 채널 방향으로 Weight를 분할하는 방식으로, Off-Chip 메모리 접근을 줄이고, 추가적인 연산을 요구하지 않는다. 제안하는 워크로드 분할 방식을 기반으로 ImageNet 분류를 위한 BCNN의 워크로드에 대해 실험한 결과, 기존의 방식과 대비, On-Chip 메모리의 요구량이 최대 89% 감소하였고, Off-Chip 메모리 접근 요구량은 최대 88% 감소하였으며, 분할 파라미터에 따른 추가적인 연산이 없다.

참고문헌 (Reference)

1 R. Andri, "YodaNN : An ultra-low power convolutional neural network accelerator based on binary weights" 236-241, 2016

2 M. Rastegari, "XNOR-Net : ImageNet classification using binary convolutional neural networks" 525-542, 2016

3 K. Simonyan, "Very deep convolutional networks for large-scale image recognition" 1-14, 2014

4 L. Jinmook, "UNPU : A 50. 6 TOPS/W unified deep neural network accelerator with 1b-to-16b fully-variable weight bit-precision" 218-220, 2018

5 H. Jin, "Software agent design with real time scheduling for embedded Systems" 443-447, 2009

6 P. Rajan, "On-chip vs. off-chip memory : the data partitioning problem in embedded processor-based systems" 5 : 682-704, 2000

7 O. Russakovsky, "ImageNet large scale visual recognition challenge" 115 (115): 211-252, 2015

8 A. Krizhevsky, "ImageNet classification with deep convolutional neural networks" 1097-1105, 2012

9 C. Szegedy, "Going deeper with convolutions" 1-9, 2015

10 Y. Umuroglu, "FINN : A framework for fast, scalable binarized neural network inference" 65-74, 2017

11 K. He, "Deep residual learning for image recognition" 1-10, 2016

12 K. Young Ho, "CASA : a convolution accelerator using skip algorithm for deep neural network" 1-5, 2019

13 I. Hubara, "Binarized neural networks" 4107-4115, 2016

14 F. Conti, "An IoT endpoint system-on-chip for secure and energy-efficient near-sensor analytics" 64 (64): 2481-2494, 2017

15 S. Dongjoo, "14. 2 DNPU : an 8. 1 TOPS/W reconfigurable CNN-RNN processor for general-purpose deep neural networks" 240-241, 2017