Biological plausibility and stochasticity in scalable VO2 active memristor neurons (1809.07867v1)

Abstract: Neuromorphic networks of artificial neurons and synapses can solve computational hard problems with energy efficiencies unattainable for von Neumann architectures. For image processing, silicon neuromorphic processors outperform graphic processing units (GPUs) in energy efficiency by a large margin, but they deliver much lower chip-scale throughput. The performance-efficiency dilemma for silicon processors may not be overcome by Moore's law scaling of complementary metal-oxide-semiconductor (CMOS) field-effect transistors. Scalable and biomimetic active memristor neurons and passive memristor synapses form a self-sufficient basis for a transistorless neural network. However, previous demonstrations of memristor neurons only showed simple integrate-and-fire (I&F) behaviors and did not reveal the rich dynamics and computational complexity of biological neurons. Here we show that neurons built with nanoscale vanadium dioxide active memristors possess all three classes of excitability and most of the known biological neuronal dynamics, and are intrinsically stochastic. With the favorable size and power scaling, there is a path toward an all-memristor neuromorphic cortical computer.

• The paper shows that VO2 memristor neurons can replicate 23 distinct spiking patterns, including three classes of neuronal excitability.
• It employs SPICE simulations to validate scalability and energy efficiency, outpacing traditional CMOS and Hodgkin-Huxley models.
• The research highlights intrinsic stochasticity that supports efficient Bayesian inference and seamless integration with existing digital architectures.

Overview of Biological Plausibility and Stochasticity in Scalable VO2 Active Memristor Neurons

This paper presents a thorough investigation into the scalability and potential biological functionality of vanadium dioxide (VO$_2$) active memristor neurons. The authors articulate the significance of this technology as an innovative alternative to traditional CMOS-based neuromorphic computing architectures, tackling both efficiency and scalability issues that have long plagued silicon-based models.

The authors delve into the architectural features and operational dynamics of VO$_2$ memristor neurons. The primary claim centers on the ability of these memristors to emulate a wide range of biological neuronal behaviors, demonstrating 23 types of spiking patterns that cover the gamut of known neuronal dynamics. Unlike previous implementations that relied heavily on overly simplistic integrate-and-fire models, this research illustrates that VO$_2$ neurons can express complex behaviors, including three classes of excitability (Class 1, Class 2, and Class 3), which are haLLMark characteristics in classifying biological neurons.

An essential aspect of these neurons lies in their intrinsic stochasticity, a critical feature for efficient Bayesian inference and other complex tasks requiring variability within neural populations. The stochastic firing and phase-locked firing (or skipping) are experimental highlights that show promise for enhanced computational models inspired by biological neural networks.

The paper extensively employs SPICE simulations to validate the potential scalability of VO$_2$ neurons in terms of dynamic and static power scaling. In simulations, memristor neurons outpace Hodgkin-Huxley model cells concerning neuron size and energy efficiency, thereby suggesting that VO$_2$ neurons could achieve biologically competitive energy efficiencies at neuron dimensions down to a few microns.

Moreover, the research introduces the concept of utilizing locally-active dynamics inherent to VO$_2$ memristors, forging a beneficial connection to the activations observed in biological volition circuits. This suggests potential for employing VO$_2$ neurons as primitive processing units within neuromorphic systems aimed at mimicking cortical computing processes.

Implications and Future Directions

The implications of this research are multifaceted. Practically, the development and integration of these memristor neurons could drastically enhance the design of energy-efficient neuromorphic hardware capable of performing complex computations with minimal energy consumption. The potential to evaporate the compatibility barriers with existing CMOS technology adds to their appeal for seamless integration into current digital infrastructures.

Theoretically, this work sets a strong precedent for further exploration into the use of non-CMOS materials and architectures in neuromorphic engineering. As industry shifts to embrace post-Moore's Law technologies, the embedded stochasticity and mimicking of biological properties by VO$_2$ neurons offer promising pathways for achieving more realistic and adaptive machine intelligence.

Future research might focus on addressing the thermal management challenges posed by the Mott transition, as well as exploring the integration of memristor-based neurons with passive memristor synapse networks for complete all-memristor neural systems. Potential developments could also investigate the geometric and material optimizations of these devices to meet diverse application needs in machine learning and real-time intelligent processing systems.

Overall, the comprehensive depiction and experimental validation of VO$_2$ active memristor neurons point toward substantial future contributions to the field of neuromorphic computing, positioning it as a viable candidate to bridge the existing performance and efficiency gap in artificial intelligence compared to biological systems.