The rapid advancement of artificial intelligence (AI) in the 21st century has led to remarkable breakthroughs across various domains, from natural language processing to computer vision. However, the exponential growth of AI is pushing traditional semiconductor technologies to their limits, demanding fundamentally new approaches to computing that can provide the necessary power and efficiency. Spintronic nanodevices, particularly Spin-Torque Nano-Oscillators (STNOs) and perpendicular Superparamagnetic Tunnel Junctions (SMTJs), have emerged as promising candidates for unconventional computing paradigms like neuromorphic computing. These devices possess unique properties such as non-volatility, analog behavior, intrinsic dynamics, and scalability, making them well- suited for emulating the key features of biological neural systems.
In the context of STNOs and SMTJs, noise manifests in various forms, including thermal fluctuations, shot noise, and electrical noise. Traditionally, these noise sources have been viewed as detrimental to device performance. However, a paradigm shift started to occur at the beginning of this thesis in the field of unconventional computing, where noise is beginning to be seen not just as an unavoidable feature, but as a potentially useful resource. This shift aligns well with the stochastic nature of biological neural systems, where it is believed that noise plays a crucial role in information processing.
In neuromorphic computing, noise can contribute to several beneficial effects. It can enhance the sensitivity of nonlinear systems to weak signals through stochastic resonance, enable the implementation of probabilistic algorithms, help break unwanted synchronization in coupled systems, and aid in escaping local minima during learning or function minimization processes. By embracing and harnessing the noise inherent in nanodevices, we open up the possibility of creating computing systems that are not only more energy-efficient but also more robust and adaptable.
This thesis investigates STNOs and SMTJs for unconventional computing, focusing mainly on their potential as basic building blocks for neuromorphic architectures. We study how these devices can function as stochastic binary neurons, harnessing their inherent noise for computation rather than trying to suppress it. By concentrating on the fundamental properties and behavior of these building blocks -from their fundamental physics to their practical implementation-, we aim to advance the development of new computing technologies. These technologies could help meet the increasing demands of AI and other computation-heavy applications while significantly reducing energy consumption.

Plus d'information :https://www.spintec.fr/phd-defense-leveraging-stochastic-properties-of-spintronic-nanodevices-for-unconventional-computing/​
​

Pour suivre la soutenance ​​​en visioconférence : https://univ-grenoble-alpes-fr.zoom.us/j/98769867024​​

​ ​