PhD: Spintronic neuromorphic computing

[ All offers ]


Location : Aarhus N, Denmark

Yearly income : https://phd.arts.au.dk/financing/salary-and-employment


Description of the offer :

Research area and project description (For application please check the following link: https://phd.tech.au.dk/for-applicants/apply-here/ Artificial neural networks are already having a very tangible effect in society. Despite the impressive progress in software, hardware is lagging behind by a decade or even more as it is not customized for implementation of such neural networks. Implementation of advanced neural network algorithms on general-purpose CPUs and GPUs is not energy-efficient. This can be overcome with the development of fully custom-designed hardware for implementation of neural networks with a vision of significantly improved energy-efficiency and performance in such a way that it resembles biological brains for specific applications such as complex signal processing and pattern recognition. Towards this goal, custom-designed CMOS implementation of artificial neurons and synapses have been researched and some progresses have been achieved. However, the CMOS implementation of neural networks is still not energy- and area-efficient [1], which will limit the scalability of such networks. These issues have driven a significant effort on investigation of non-CMOS implementations of neuromorphic computing systems, which include synapses implemented using memristors [2,3], Magnetic Tunnelling Junctions (MTJs) etc., and artificial neurons using MTJs, Spin-Torque Nano-Oscillators (STNOs) [4], mutually synchronized Spin-Hall Nano-Oscillators (SHNOs) [5], etc.

Among different technologies, Spintronics is a strong contender as it is compatible with CMOS, multifunctional, extremely versatile with features like non-volatility, plasticity and oscillatory behaviour, which can be exploited to implement artificial neural components to develop energy-efficient neural networks. Despite some progress in spintronic NCS [4, 6], challenges such as lack of proper artificial neuron and synapse implementation resembling biological neural components, low speed, lack of any forward-looking architecture implementation, lack of proper algorithms for spintronic networks, lack of any simulation platforms that combine libraries from several emerging technologies together for a better assessment of larger networks, have hindered scaling of complex architectures [7].

Within an interdisciplinary project, SpinAge, funded by H2020 [8], we aim at development of a brain-inspired (neuromorphic computing) networks using synchronized spintronic oscillators [9] together with other European research teams [8] leading to orders of magnitudes improvement in performance. For this position, we aim at development of analytical and numerical models for spintronic devices (oscillators) [10-12] and bringing them into a simulation platform enabling simulation of larger neural networks, which requires a good understanding of the physics behind the spintronic oscillators, and skills in development of analytical models for such devices.

[1] J. Kim et al., Proc. of the IEEE, 103 (1), 106-130, (2015).
[2] S. Ambrogio et al., Nanotechnology 24, 384012, (2013).
[3] S. Saïghi et al., Front. Neurosci. 9:51, (2015).
[4] J. Torrejon, , et al., Nature 547.7664, 428-431, (2017)
[5] A. Awad et al., Nature Phys. 13, 292, (2017)
[6] H. Farkhani, et al. Frontiers in Neuroscience 13 (2019).
[7] Grollier, J., et al. Nature Electronics (2020): 1-11.
[8] https://cordis.europa.eu/project/id/899559
[9] M. Zahedinejad, et al. Nature Nanotechnology 15.1, 47-52, (2020).
[10] S. G. Ramasubramanian, et al. ISLPED. IEEE, (2014).
[11] D. I. Albertsson, et al. IEEE Transactions on Magnetics 55.10, 1-8, (2019).
[12] J. Pelloux-Prayer, and Farshad Moradi. IEEE Transactions on Electron Devices (2020).


[ All offers ]