Frank Mizrahi presents a tutorial at the 7th IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS 2025) in Bordeaux. Key message ? “AI hardware with Spintronics is a high-reward challenge for circuit designers and architects“
Julie Grollier will give a lecture on emotional imprints and poetry in the language of physics at the upcoming Gravitations workshop, taking place June 12–13 in Strasbourg. This talk marks a first attempt to bring together two of her long-standing interests and explore the connections between them
The amazing Danijela Marković explained everything about AI and neuromorphic quantum computing last week at the French-Serbia innovation forum. Classy! https://sfif.rs/
We propose to implement a recurrent neural network in hardware using spintronic oscillators as dynamical neurons. Using numerical simulations, we build a multilayer network and demonstrate that we can use back-propagation through time and standard machine-learning tools to train this network. The wide range of time scales that the network can handle enables high accuracy on time-series classification.
https://link.aps.org/doi/10.1103/PhysRevApplied.23.034051
https://arxiv.org/abs/2408.02835
Our group meetings are very interdisciplinary. Some of us focus on machine learning, working at their desk mostly, while others favor sample fabrication in the clean room. So we try to be as pedagogical as we can. Here Baptiste trains for his talk on Bosonic reservoir computing.
Baptiste obtained very cool experimental results, showing how to induce and exploit non-linearity in quantum reservoirs. He will present his work at this workshop in Denmark.
Article under preparation!
Very nice feature interview of Julie on the website of Université Paris-Saclay ! On this topic, Julie says: “I am trying to share more of my true self, my emotions and core motivations for physics in interviews recently. I think it is important in order to promote diversity and novel avenues in research.”
Congratulations to Hanuman Singh for obtaining a senior research scientist position in VTT in Finland on quantum materials and sensors. We are happy for him, and sad that he is leaving, he did so much for the group. In less than 18 months, he obtained really beautiful results on novel synaptic devices. A few more steps to unlock and we will be able to share them with you!
We have used the properties of topologically protected magnetic particle-like structures called skyrmions to perform a fundamental operation of neuromorphic computing: the weighted summation of synaptic signals. These skyrmions, which act as analogs of neurotransmitters in a biological neural network, enabled the reproduction of this operation in a compact and energy-efficient manner, opening up new possibilities for neuromorphic components that approach the efficiency of biological systems. The recently published article is here: Nature Electronics 2025 [arXiv]
Figure: Weighted summation in a device composed of two parallel tracks (synapses) made of a magnetic multilayer.
(a-d) Kerr microscopy images of the device, which consists of two parallel magnetic multilayer tracks, each 6 µm wide, connected by a transverse Ta Hall electrode, also 6 µm wide. After magnetization saturation of the track (a), skyrmions can be selectively nucleated in track 1 (b) and track 2 (c), before being erased by an inverse field or current (d).
(e) Hall voltage ∆V (in red) and the corresponding sum of the detected skyrmion count in both tracks, ∑NSk, detec (in blue), during the successive injection of skyrmions into the tracks. 20 nucleation pulses are sequentially applied to each track (indicated by the green and yellow regions for tracks 1 and 2, respectively) using current pulses of approximately 116 GA/m² and 50 ns, at µ0Hz = 20 mT. The thin red curve represents the raw electrical measurements after drift correction, while the thick curve is the same after smoothing.
Our article, Classification of multi-frequency RF signals by extreme learning, using magnetic tunnel junctions as neurons and synapses by Nathan Leroux et al., is featured in the “2023 Papers with Best Practices in Data Sharing and Comprehensive Background Review” list of APL Machine Learning.
You can also read a nice Scilight (Editor’s highlight) about our work: Using spintronic devices in neural networks to analyze radio signals
Dongshu showed that learning algorithms based on error backpropagation are compatible with unsupervised learning. All you have to do is to define an “unsupervised target”. Preprint coming soon!
