Session Info
Hardware Keynote: The Multifaceted Impact of Memristors on Neuromorphic Systems
The Multifaceted Impact of Memristors on Neuromorphic Systems
Abstract
Memristors, also known as resistive memories, are nanodevices that mimic artificial synapses and possess the potential to revolutionize neuromorphic design. This talk explores the transformative capabilities of memristors by examining their application in neuromorphic systems. While these technologies have matured sufficiently to enable the fabrication of complete systems, they present challenges due to their atomic-scale features, leading to a high degree of variability. Through examples of fabricated hybrid CMOS/memristor circuits, we demonstrate various techniques that allow neuromorphic systems to harness the benefits of memristors while mitigating their drawbacks. In the realm of analog neuromorphic systems, memristors serve as in-memory computational units. Their unpredictable behavior necessitates the use of dedicated programming strategies to achieve energy-efficient artificial intelligence. To illustrate this, we showcase a neural network for arrhythmia detection that maintains stable accuracy over an extended period of two months. Furthermore, we emphasize the potential of memristors in digital systems, which can yield low-power and highly robust systems capable of self-powered operation. We present a Bayesian machine that demonstrates near immunity to single-event upsets. Additionally, we describe a fabricated digital neural network consisting of 32,768 memristors powered by a miniature solar cell. This system seamlessly switches between exact and approximate computation depending on the available energy. Lastly, we delve into the realm of stochasticity, where memristor variability can be transformed into an advantage, driving superior efficiency and robustness. We showcase how this variability can enable neural networks to function as Bayesian neural networks capable of estimating prediction uncertainty. Furthermore, we demonstrate that memristor variability facilitates a form of Bayesian learning, particularly effective in situations with limited data.
