Memristive Neurons and Synapses for Neuromorphic Edge Computing

Introduction

In recent years, Artificial Intelligence has shifted towards collaborative learning paradigms, where multiple systems acquire and elaborate data in real-time and share their experience to improve their performance. MEMRINESS will generate new fundamental computing primitives that will overcome the current challenges for the deployment of intelligent systems on the edge.

System Requirements

The requirements of a system operating on the edge are very tight:

• Power efficiency
• Low area occupation
• Fast response times
• Online learning

Brain-inspired architectures such as Spiking Neural Networks (SNNs) use artificial neurons and synapses that perform low-latency computation and internal-state storage simultaneously with very low power consumption. However, they mainly rely on standard technologies, which make SNNs unfit to meet the above-mentioned constraints.

Challenges in Current Technologies

Indeed, the dream of compact and efficient neurons and synapses, able to work at different time scales to match real-time constants and to retain memory of their state even in the absence of a power supply, cannot be realized without flanking standard technologies with emerging ones.

Promising Solutions

In this respect, memristive technology has shown promising results due to its ability to support non-volatile storage of the SNN parameters. Yet so far, research has prioritized the non-volatile properties of the devices rather than focusing additionally on the reproduction of multi-temporal synaptic and neural dynamics.

Development Approach

To solve this problem, I will develop neurons and synapses that exploit the intrinsic physical characteristics and dynamics of volatile and non-volatile memristive devices to enable the design of compact, power-efficient SNNs with multi-timescale dynamics.

I will use a holistic approach and co-develop every aspect, from the devices to the circuits, to the learning algorithms.

Demonstration of Capabilities

I will use the results to design an SNN and demonstrate its collaborative and online learning capabilities in three scenarios of increasing complexity.