Ultra-low power neuromorphic devices based on the manipulation of topological spin textures – NeuroSky

The digital and electric transformation of our society is enabled by high-performance integrated circuits which require to solve increasingly complex computing tasks. The development of artificial neural networks in the last decade enabled a massive performance gain in computing hardware. However, this progress has led to a large rise in power consumption due to poorly optimized hardware architectures and computationally demanding training algorithm that require cloud computing. Besides the vast energy cost, cloud computing also poses serious concerns for customer privacy and the vulnerability of the system to hacking. To address this issue, there is a need for novel compact, integrable, low energy hardware that allows local, in-chip training and fast real-time inference (edge AI) in autonomous and embedded systems.
NeuroSky propose a novel ultra-low energy, fast, and compact hardware solution to solve intensive and complex cognitive tasks.
Our approach is based on reservoir and deep physical neural network computing, which exploits the ultra-low power excitation of nanoscale disordered topological spin textures to perform high performance cognitive tasks. The topological spin textures will be integrated in a magnetic tunnel junction (MTJ) to excite its dynamics electrically with low power and detect it with a large electrical readout signal. This approach combines the advantages of spintronic devices while allowing complex recognition tasks with very low energy by exploiting the spin texture intrinsic memory, complexity, non-linearity and their low power excitation. The objective of the project is to demonstrate the proof of concept of devices based on this approach, evaluate their performance and benchmark it to existing hardware implementations. We will also explore spatio-temporal coupled reservoirs such as coupled MTJs and deep physical neural network implementation and evaluate the scalability to larger arrays. To reach this objective, we will develop a large panel of novel and unexplored scientific and technical solutions covering the field of material engineering, device fabrication and characterization, metrology, and algorithmic and hardware design.
The development of a ultralow power edge AI hardware solution proposed in Neurosky addresses the current broad societal need for more energy efficient AI chips with increased data protection and privacy which operates without any cloud connection. The technology could have a signficant societal and economical impact on a broad range of applications ranging from smart sensors in space and harsch environment, predictive maintenance, healthcare and automotive.