[Paper Review] Implementation of Multilayer Perceptron Network with Highly Uniform Passive Memristive Crossbar Circuits

F. Merrikh Bayat, M. Prezioso+3 more|arXiv (Cornell University)|Dec 4, 2017

Advanced Memory and Neural Computing2 references18 citations

TL;DR

This paper demonstrates a fully hardware-implemented multilayer perceptron classifier using two 20×20 passive memristive crossbar arrays integrated with CMOS components, achieving 97% of the simulation accuracy through improved memristor uniformity. The system achieves high classification fidelity with an ex-situ training approach, marking a significant advance in neuromorphic computing with analog memristive hardware.

ABSTRACT

The progress in the field of neural computation hinges on the use of hardware more efficient than the conventional microprocessors. Recent works have shown that mixed-signal integrated memristive circuits, especially their passive ('0T1R') variety, may increase the neuromorphic network performance dramatically, leaving far behind their digital counterparts. The major obstacle, however, is relatively immature memristor technology so that only limited functionality has been demonstrated to date. Here we experimentally demonstrate operation of one-hidden layer perceptron classifier entirely in the mixed-signal integrated hardware, comprised of two passive 20x20 metal-oxide memristive crossbar arrays, board-integrated with discrete CMOS components. The demonstrated multilayer perceptron network, whose complexity is almost 10x higher as compared to previously reported functional neuromorphic classifiers based on passive memristive circuits, achieves classification fidelity within 3 percent of that obtained in simulations, when using ex-situ training approach. The successful demonstration was facilitated by improvements in fabrication technology of memristors, specifically by lowering variations in their I-V characteristics.

Motivation & Objective

• To develop a fully hardware-implemented multilayer perceptron classifier using passive memristive crossbars for efficient neuromorphic computing.
• To overcome the limitations of immature memristor technology by improving device uniformity and reducing I-V characteristic variations.
• To demonstrate a scalable, mixed-signal neuromorphic system with higher complexity than previous passive memristive classifiers.
• To validate the system's performance using an ex-situ training approach and compare it to simulation benchmarks.
• To enable practical deployment of memristor-based neural networks by integrating passive memristive arrays with discrete CMOS components.

Proposed method

• The system uses two 20×20 passive memristive crossbar arrays to implement the weight matrices of a one-hidden-layer perceptron.
• Memristors are fabricated with improved uniformity to reduce variations in their I-V characteristics, enhancing device consistency.
• The crossbar arrays are board-integrated with discrete CMOS components for signal conditioning, biasing, and readout.
• An ex-situ training approach is employed, where network weights are computed offline and mapped to the memristive crossbars.
• The system performs inference using mixed-signal processing, with analog weights and digital control logic.
• Classification performance is evaluated by comparing hardware results to simulation benchmarks.

Experimental results

Research questions

• RQ1 Can a fully hardware-implemented multilayer perceptron be realized using passive memristive crossbars with high fidelity?
• RQ2 To what extent does improved memristor uniformity enhance the performance of neuromorphic hardware?
• RQ3 How does the classification accuracy of a hardware-based multilayer perceptron compare to simulation when using ex-situ training?
• RQ4 Can passive memristive crossbars support a network complexity significantly higher than previously demonstrated?
• RQ5 What is the role of CMOS integration in enabling stable and accurate operation of memristive neuromorphic systems?

Key findings

• The hardware-implemented multilayer perceptron achieved classification accuracy within 3% of the simulation benchmark, demonstrating high fidelity.
• The system's complexity is nearly 10 times higher than previously reported functional neuromorphic classifiers based on passive memristive circuits.
• Improved memristor fabrication reduced I-V characteristic variations, enabling reliable operation of the crossbar arrays.
• The integration of passive memristive crossbars with discrete CMOS components enabled stable and accurate inference.
• The ex-situ training approach successfully mapped precomputed weights to the hardware, ensuring consistent performance.
• The results confirm the feasibility of using passive memristive crossbars for scalable, high-performance neuromorphic computing.