Meta-Learning of memristive devices for lifelong adaptation

ongoing

The continued growth of artificial intelligence (AI) in the cloud is driving up global energy costs. As a result, a paradigm shift is taking place where new intelligent devices are placed right at the edge. MALEFICENT will create a new framework for implementing sustainable AI at the edge using standard and novel technologies. Neuromorphic systems using emerging memory devices such as resistive switching devices (ReRAM) or ferroelectric capacitors (FeCap), are a promising alternative for AI systems thanks to their energy efficiency and non-volatility. However, the deployment of these devices in real-world applications poses some challenges, due to their intrinsic variability and limited bit precision. I will use advanced learning techniques such as meta-learning to create a self-adaptive neuromorphic system based on emerging memory devices able to exploit the intrinsic features of the devices while mitigating their limitations. I will then apply it to a real-world environment, such as robotics. This will have a significant impact on the research of emerging memory technologies, by opening up the possibility of exploitation in an industrial context.