SurfaceScience21 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
O: Fachverband Oberflächenphysik
O 68: Poster Session V: Poster to Mini-Symposium: Manipulation and control of spins on functional surfaces I
O 68.3: Poster
Mittwoch, 3. März 2021, 10:30–12:30, P
An atomic Boltzmann machine capable of self-adaption — Brian Kiraly1, •Elze J Knol1, Werner MJ van Weerdenburg1, Hilbert J Kappen2, and Alexander A Khajetoorians1 — 1Institute for Molecules and Materials, Radboud University, Nijmegen, the Netherlands — 2Donders Institute, Radboud University, Nijmegen, the Netherlands
To move beyond the current hybrid approaches to hardware-based artificial neural networks, new architectures, linking physical phenomena to machine learning models, are needed. Here, we realized an atomic Boltzmann machine capable of self-adaption using atomic manipulation with a scanning tunneling microscope. We utilized the concept of orbital memory, derived from single Co atoms on black phosphorus, as the building blocks of the prerequisite multi-well energy landscape. Namely, when gating two Co atoms simultaneously, there is a finite probability in each of the four possible states. This multi-well behavior persists for larger ensembles. Additionally, we found that the coupling between Co atoms is anisotropic, which we exploited to build synapses capable of tuning the neurons' energy landscape, and to introduce two inherent timescales: a fast neural timescale and a slow synaptic timescale. Finally, we observed self-adaption of the synaptic weights in response to external electrical stimuli, opening a path to on-chip learning in atomic-scale machine learning hardware.
B. Kiraly et al., arXiv:2005.01547v2 (2020)