Speaker
Descrizione
The development of neuromorphic devices is a key step toward enabling low-power artificial intelligence. Among the various emerging technologies, memristive devices have attracted considerable attention owing to their ability to perform efficient one-shot multiply-accumulate (MAC) operations, which are fundamental for neural network computation. However, programming noise remains a major challenge, as it significantly degrades network performance and is intrinsically associated with the operation of memristive devices.
A promising solution is offered by a class of molecular spin valves. Here, we investigate the synaptic behavior of molecular La₀.₇Sr₀.₃MnO₃/tris(8-hydroxyquinolinato)gallium/AlOₓ/Co spintronic devices, which uniquely combine memristive and magnetoresistive functionalities [1,2,3,4]. By arranging these devices in an N-crosspoint architecture and encoding the synaptic weight in the total conductance, we realize an artificial synapse that is robust against programming noise [5].
This robustness is achieved through a quantization strategy in which each crosspoint represents a bit of different significance. The memristive behavior is first exploited to program the synaptic weight, while the magnetic degree of freedom is subsequently used to select the conductance configuration that minimizes programming noise.
To validate the proposed architecture, we experimentally demonstrated the quantization strategy and implemented a reinforcement learning task on a simple neural network. The performance of the quantized network was compared with that of a conventional architecture employing the same number of crosspoints under progressively increasing programming noise. The proposed synapse consistently exhibited superior robustness, maintaining higher performance under noisy operating conditions.
These results highlight the strong potential of molecular resistive spin valves as fundamental building blocks for next-generation neuromorphic architectures, offering an effective hardware strategy to improve the reliability of memristive neural networks.
⦋1⦌ Shumilin A., Baldassini C., et al. Glassy Synaptic Time Dynamics in Molecular La₀.₇Sr₀.₃MnO₃/Gaq₃/AlOₓ/Co Spintronic Crossbar Devices. Advanced Electronic Materials. 2024; 10(2): 2300887. DOI: 10.1002/aelm.202300887.
⦋2⦌ Riminucci A., Legenstein R. Fast Learning Synapses with Molecular Spin Valves via Selective Magnetic Potentiation. arXiv preprint. 2019. DOI: 10.48550/arXiv.1903.08624.
⦋3⦌ Prezioso M., Riminucci A., Graziosi P., et al. A Single-Device Universal Logic Gate Based on a Magnetically Enhanced Memristor. Advanced Materials. 2013; 25(4): 534–538. DOI: 10.1002/adma.201202624.
⦋4⦌ Baldassini C. Spintronic Advantage in the Training of a Molecular Cross-Bar Neural Network. Unpublished manuscript. 2025.
⦋5⦌ Baldassini C., Riminucci A., et al. Programming Noise Mitigation by Adaptive Quantization in Spintronic Memristive Artificial Synapses. Submitted for publication.
| Giovane Ricercatore (under 40) | Yes |
|---|
