Speaker
Descrizione
Increasing the computing and memory efficiency of artificial intelligence is a major driver in current research. In memory computing tackles this problem at the hardware level by removing the memory-computing bottleneck intrinsic to current von Neumann architectures. At the algorithmic level, approaches such as multi-experts or low rank adaptation attempt to achieve the same performance with less memory.
In this work we combine these approaches in a novel, hardware memristive-spintronic approach[1]. We use the memristive properties to implement in memory computing, while we use the spintronic ones to switch between different, pre-stored experts.
We tested this approach in simulations of reinforcement learning which together with supervised and unsupervised learning, is one the three paradigms for learning in artificial intelligence. Reinforcement learning is based of rewarding the system when the desired goal is achieved and is a primary tool for robotics. The simulations were aimed at demonstrating the ability to train several experts simultaneously on the same physical substrate. We used inverted pendulum balancing, a classic reinforcement learning benchmark, to test our concept, and demonstrated the ability to train the hardware to balance three distinct inverted pendula[2], [3].
Bibliography
[1] A. Shumilin et al., «Glassy Synaptic Time Dynamics in Molecular La0.7Sr0.3MnO3/Gaq3/AlOx/Co Spintronic Crossbar Devices», Advanced Electronic Materials, vol. 10, fasc. 8, p. 2300887, 2024, doi: 10.1002/aelm.202300887.
[2] C. Baldassini et al., «Spintronic Advantage of Molecular Spin-Valves for Reinforcement Learning», in 2026 IEEE International Magnetic Conference - Short Papers (INTERMAG Short Papers), apr. 2026, pp. 1–2. doi: 10.1109/INTERMAGShortPapers68882.2026.11596177.
[3] C. Baldassini et al., «Multiweight molecular spintronic synapses for reinforcement learning», in Spintronics and Nanomagnetism, SPIE, mag. 2026, p. 38. doi: 10.1117/12.3100181.
| Giovane Ricercatore (under 40) | No |
|---|
