Within the ERC Proof of Concept ‘SOFT’ we made a surprising discovery: the magnetic Domain Wall (DW) motion over energy barriers is in many aspects similar to the electric charging, discharging and firing process of a biological spiking neuron. Our bio-mimicking neuron is far superior to the artificial neurons composing the state-of-the-art artificial neural networks. We demonstrated experimentally that a very simple network of such spiking neurons is able to perform simultaneously two different tasks (speech recognition and speaker identification) without the need of any prior feature extraction.
When subjected to magnetic field pulses, the DW climbs the energy barrier, but relaxes back towards the bottom of the well in the inter-pulse period. Based on this principle, we have designed, fabricated and tested a new type of magnetic spiking neuron that combines the complexity of biological neurons with the stability and scalability of standard technological methods. Our magnetic neuron shares many features with the biological neurons (real-time operation, stochasticity, adaptation, saddle-node bifurcation, refractory period), while being much more robust and reliable than its biological counterpart. We have shown that the magnetic neurons are truly bio-mimicking, by testing the ability of the magnetic neural network to perform a bio-mimicking task. We converted human voice into spikes using zero-crossings, a mechanism known to be present in the mammalian ear, and tested the ability of the system to perform pattern recognition. Similarly to the auditory nervous system, the magnetic neural network, works in real-time (processing on-the-fly, without signal compression), is task agnostic (able to simultaneously recognize the speech and identify the speaker), and able to generalize very well from only few examples (90% recognition rate, with only 10% of the data used for training).
To the best of our knowledge, this is the only technology that supports bio-mimicking neurons working in real-time. Next, our goal is to imitate as much as possible the natural processes enabling human skills. Although the natural intelligence is far superior to our technology, our success with speech recognition and speaker identification gives us confidence that the ability to recognize patterns in time-varying signals is within reach. Our next goal is to create a prototype for a device that uses magnetic neural networks to ‘listen’ to machines in a manufacturing line, detecting small changes in the pattern of analog signals (sound; pressure; voltage; temperature; etc…) that precede a future malfunction.
We aim to validate and implement our novel neural network for an emerging market with a significant unmet need. Our project’s focus is on a single problem that is perfectly suitable for the present development of our technology, represents a major societal challenge and is a perfect economic opportunity: predictive maintenance and quality control in manufacturing facilities. Losses due to failing equipment and unstable fabrication processes affect significantly the cost of goods and as well as their environmental impact. The prevalence of these problems is reflected by the market size and growth (CAGR=30%; €60B by 2030). This application is perfectly suitable for the present state of our technology: there are no constrains in terms of miniaturization, energy autonomy, remote connectivity; The size of our device is much smaller than the equipment generally present in a manufacturing facility. The device can be plugged for power and it can use standard short distance connections if needed. Moreover, for many industries, data privacy is critical.
Within the ERC Proof of Concept project SOFT we performed the early exploration of the market by testing potential demand and acceptability with end users. We presented our innovation to a company based in Romania that manufactures paper tissues (yearly revenue of €100 million, €16 million profit) and a second company based in Romania involved in metal works (undisclosed financials, approximately 100 employees). They validated the problem (unforeseen breakdowns of equipment that induce manufacturing delays and increase in costs) and acknowledged the lack of suitable technologies that they could implement with a view towards predictive maintenance.
A second type of potential customer is the industrial equipment manufacturer. Our solution can also be integrated directly into new equipment. We have initiated contacts with two suppliers for the semiconductor industry. The first is a SME with 20 employees which produces testing equipment for Magnetic Random Access Memories. Their clients include IMEC, CEA, TSMC and Samsung. The second is the French division of a large company (27000 employees) with whom we plan to initiate a collaboration. Their interest in our solution is motivated by the high cost of downtime in a semiconductor fabrication/testing line, and the reticence of the semiconductor foundries to use cloud-based solutions, for obvious privacy issues. Integrating a “privacy-guaranteed” AI hardware module into their equipment offers an important commercial and marketing advantage over their competitors.
The ERC Proof of Concept ‘SOFT’ led to the incorporation of a new spin-off (Golana Computing), which will be tasked with the further development of the technology towards a marketable product.
