[Paper Review] Binary Weighted Memristive Analog Deep Neural Network for Near-Sensor Edge Processing

Olga Krestinskaya, Alex Pappachen James|arXiv (Cornell University)|Aug 2, 2018

Advanced Memory and Neural Computing12 references4 citations

TL;DR

This paper proposes a binary-weighted memristive analog deep neural network using binary-state memristive devices to enable low-power, area-efficient near-sensor edge inference. By leveraging binary weight updates via backpropagation in a crossbar architecture, the network achieves ~90% accuracy on MNIST digit recognition, demonstrating feasibility for energy-efficient edge AI with minimal hardware overhead.

ABSTRACT

The memristive crossbar aims to implement analog weighted neural network, however, the realistic implementation of such crossbar arrays is not possible due to limited switching states of memristive devices. In this work, we propose the design of an analog deep neural network with binary weight update through backpropagation algorithm using binary state memristive devices. We show that such networks can be successfully used for image processing task and has the advantage of lower power consumption and small on-chip area in comparison with digital counterparts. The proposed network was benchmarked for MNIST handwritten digits recognition achieving an accuracy of approximately 90%.

Motivation & Objective

• To address the challenge of limited switching states in memristive devices for analog neural network implementation.
• To enable energy-efficient, on-chip inference for edge computing by using binary-weighted memristive devices.
• To design a trainable analog DNN using backpropagation with binary weights on memristive crossbars.
• To benchmark performance on image classification tasks with minimal hardware footprint.
• To demonstrate feasibility of near-sensor processing using analog, binary-weighted memristive networks.

Proposed method

• Utilizes binary-state memristive devices to represent weights in a crossbar array, enabling analog computation.
• Applies a modified backpropagation algorithm to update weights in binary form (±1) during training.
• Employs a hybrid analog-digital architecture where weights are stored in memristors and activations are processed in analog domain.
• Designs a network with binary-weighted layers to reduce power and area compared to full-precision digital implementations.
• Optimizes the training process to maintain accuracy despite binary weight constraints.
• Validates the architecture on MNIST using a fully connected deep neural network with binary weights.

Experimental results

Research questions

• RQ1 Can binary-weighted memristive crossbars effectively implement trainable analog deep neural networks?
• RQ2 How does the accuracy of a binary-weighted memristive DNN compare to full-precision digital counterparts on image tasks?
• RQ3 What are the power and area advantages of using binary-state memristors in edge AI hardware?
• RQ4 Can backpropagation be effectively adapted to binary weight updates in a memristive crossbar architecture?
• RQ5 To what extent can analog computation with binary weights enable near-sensor edge inference?

Key findings

• The proposed binary-weighted memristive DNN achieves approximately 90% accuracy on the MNIST handwritten digit recognition benchmark.
• The network demonstrates significant power and area efficiency compared to digital implementations due to binary weight representation and analog in-memory computation.
• The use of binary-state memristors enables practical implementation despite limited switching states in real devices.
• The architecture supports end-to-end training with backpropagation, validating the feasibility of learning in a binary-weighted analog framework.
• The design enables near-sensor processing by minimizing data movement and on-chip memory requirements.
• The results confirm that binary-weighted analog networks can achieve acceptable accuracy with low hardware cost for edge AI applications.