Spintronic Spiking Neurons – SpinSpike

The recent progress of deep neural networks has enabled artificial intelligence (AI) to disrupt many fields, from scientific research to commercial applications. One major concern however is the energy cost of AI, that poses environmental threats and prevents its deployment in embedded systems. Spiking neural networks are among the most promising algorithms for implementing low power hardware capable of real-time learning. Unlocking their potential for energy efficient operation requires to emulate synapses and spiking neurons with physical nanodevices and connect them closely and densely as in the brain. Spintronics is a promising technology for the implementation of such networks due to its low energy consumption and very high endurance that could enable life-long learning. Yet, one key ingredient is missing to build spintronic spiking neuromorphic systems: the spintronic spiking neuron itself.
The global goal of SpinSpike project is to pave the way for an all-spintronic low power spiking neural networks using magnetic tunnel junctions as nanoscale spiking neurons and synapses. To reach this goal we have defined three ambitious objectives:
1) We will demonstrate the spiking functionality of individual magnetic tunnel junction neurons. We have identified two promising magnetic tunnel junction configurations with different dynamic transient spiking and demonstrated their feasibility through preliminary micromagnetic simulations and experiments. The first concept uses a perpendicular polarizer that induces a single period of magnetization oscillation in a planar free layer when a current pulse is applied, that is then converted into a voltage spike through magnetoresistance. The second concept exploits the windmill motion of magnetizations in an all-perpendicular magnetic tunnel junction with two weakly coupled free layers to generate transient voltage spikes.
2) We will demonstrate with magnetic tunnel junctions two key functionalities of spiking neural networks: STDP and synchronization. One of the most promising unsupervised learning rule – inspired from biological results – is Spike Timing Dependant Plasticity (STDP). However, a demonstration of STDP induced by spiking nano-neurons on nano-synapses has yet to be achieved. In SpinSpike, we will achieve the first experimental demonstration of STDP using two spiking MTJ neurons and one MTJ synapse. The spikes emitted by biological neurons can synchronize, giving rise to oscillations that can in turn modify the dynamics of distant neural assemblies, which is a key process in neuroscience for computing and memory. In SpinSpike we will demonstrate the synchronization between spikes of two MTJ neurons and explore how to control this synchronization through an intermediate MTJ synapse.
3) We will evaluate the capabilities of spintronic spiking neural networks for unsupervised learning applications. Implementing a fully spintronic spiking neural network requires a co-design of the hardware and the algorithms. Through large-scale simulations based on the developed device physical models, we will reinvent spiking algorithms to unlock the benefits of spintronic based neurons and synapses with the final goal to evaluate spintronic spiking neural networks for unsupervised learning applications.
By achieving these three objectives, SpinSpike will harness the multi-functionality, endurance and stability of nanoscale spintronic devices to provide novel, low power approaches for building large scale spiking neural networks that learn to recognize input data in an unsupervised and energy-efficient way. This will be achieved by a multidisciplinary approach leveraging the expertise of the partners in spintronics, neuromorphic computing and AI, through a close-loop interaction between design optimization through micromagnetic simulations, nanofabrication and electrical testing.