Memristive Circuit Design of Quantized Convolutional Auto-Encoder

This work proposes a memristor-based quantized convolutional auto-encoder (MQCAE) and applies it in an image denoising application. The pulse width or amplitude is gradually tuned by incremental steps in current memristive programming methods, which can be extremely time-consuming. In this work, a m...

Full description

Saved in:
Bibliographic Details
Published in:IEEE transactions on emerging topics in computational intelligence 2023-08, Vol.7 (4), p.1301-1313
Main Authors: Zhang, Yaozhong, Wang, Xiaoping, Yang, Chao, Zeng, Zhigang
Format: Article
Language:English
Subjects:
Acceleration
Analog circuits
Analog data
Circuit design
Coders
Convolution
convolutional auto-encoder
crossbar array
Deconvolution
image denoising
Kernel
memristor
Memristors
Modules
Noise reduction
Pipelines
Programming
Pulse duration
quantization
Table lookup
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:This work proposes a memristor-based quantized convolutional auto-encoder (MQCAE) and applies it in an image denoising application. The pulse width or amplitude is gradually tuned by incremental steps in current memristive programming methods, which can be extremely time-consuming. In this work, a multi-level programming method without incremental steps is proposed. MQCAE is composed of five kinds of network layers: convolution layers, deconvolution layers, activation function layers, batch normalization layers and max-pooling layers. We design a memristive circuit for realizing convolution and deconvolution with different kernel parameters. The proposed circuit generates one output feature row every cycle. Analog data buffers are designed to store the intermediate data among network layers. In addition, a reconfigurable analog circuit for realizing activation functions and batch normalization is presented. By using analog temp modules, convolution layers and max-pooling layers are able to compute simultaneously. We construct the MQCAE and introduce the pipeline technique among network layers based on the circuit modules. As a result, MQCAE is able to process one FashionMNIST image in 36 cycles with clock frequency 20 kHz. Finally, we verify the effectiveness of MQCAE in an image denoising task. The results show that the denoising performance of proposed scheme is close to software model while the processing speed is faster.
ISSN:2471-285X
2471-285X
DOI:10.1109/TETCI.2022.3214582