Summary
Golana Computing is a new start-up company, spin-off from Spintec-CNRS, exploiting a recent scientific and technological breakthrough in the design and fabrication of bio-mimicking magnetic neurons.
Breakthrough: Studying new magnetization reversal schemes, we have inadvertently discovered that domain wall depinning from geometrical traps imitates in many regards the spiking of biological neurons. Based on this, we designed, fabricated and tested magnetic neurons able to complete bio-mimicking tasks. Our magnetic neural network performs speech recognition and speaker identification in real-time, without any prior feature extraction. The audio is simply transformed into spikes by a mechanism inspired from the mammalian ear.
Our Goal is to develop a technology and fabricate a prototype that extends this unique ability to other types of analog signals, and apply it for predictive maintenance in the manufacturing industry. The present solutions based on mainstream artificial intelligence (AI) struggle, because the problems at hand are too fragmented: the training data is too scarce and the model engineering relies on very specific expert knowledge.
Our Solution, frugal in terms of data and resources, based on a “task-agnostic” generic device, is able to identify unusual patterns in the analog signals. Our bio-mimicking approach should imitate the ability of human technicians, which assess the state of their machines by the sound. On the long term, our technology could be adapted for a variety of AI applications requiring low energy consumption or full privacy.
The EIC Transition call corresponds exactly to our present needs: accelerate the development and the market readiness of our technology. Moreover, we address explicitly the requirements for Green Digital Devices. By working “on the edge”, our device reduces the energy and resources required for data transfer and, by imitating the biological neurons, it also reduces the energy required for the computation.
Breakthrough: Studying new magnetization reversal schemes, we have inadvertently discovered that domain wall depinning from geometrical traps imitates in many regards the spiking of biological neurons. Based on this, we designed, fabricated and tested magnetic neurons able to complete bio-mimicking tasks. Our magnetic neural network performs speech recognition and speaker identification in real-time, without any prior feature extraction. The audio is simply transformed into spikes by a mechanism inspired from the mammalian ear.
Our Goal is to develop a technology and fabricate a prototype that extends this unique ability to other types of analog signals, and apply it for predictive maintenance in the manufacturing industry. The present solutions based on mainstream artificial intelligence (AI) struggle, because the problems at hand are too fragmented: the training data is too scarce and the model engineering relies on very specific expert knowledge.
Our Solution, frugal in terms of data and resources, based on a “task-agnostic” generic device, is able to identify unusual patterns in the analog signals. Our bio-mimicking approach should imitate the ability of human technicians, which assess the state of their machines by the sound. On the long term, our technology could be adapted for a variety of AI applications requiring low energy consumption or full privacy.
The EIC Transition call corresponds exactly to our present needs: accelerate the development and the market readiness of our technology. Moreover, we address explicitly the requirements for Green Digital Devices. By working “on the edge”, our device reduces the energy and resources required for data transfer and, by imitating the biological neurons, it also reduces the energy required for the computation.
Unfold all
/
Fold all
More information & hyperlinks
| Web resources: | https://cordis.europa.eu/project/id/101113180 |
| Start date: | 01-05-2023 |
| End date: | 30-04-2026 |
| Total budget - Public funding: | 2 499 999,00 Euro - 2 499 999,00 Euro |
Cordis data
Original description
Golana Computing is a new start-up company, spin-off from Spintec-CNRS, exploiting a recent scientific and technological breakthrough in the design and fabrication of bio-mimicking magnetic neurons.Breakthrough: Studying new magnetization reversal schemes, we have inadvertently discovered that domain wall depinning from geometrical traps imitates in many regards the spiking of biological neurons. Based on this, we designed, fabricated and tested magnetic neurons able to complete bio-mimicking tasks. Our magnetic neural network performs speech recognition and speaker identification in real-time, without any prior feature extraction. The audio is simply transformed into spikes by a mechanism inspired from the mammalian ear.
Our Goal is to develop a technology and fabricate a prototype that extends this unique ability to other types of analog signals, and apply it for predictive maintenance in the manufacturing industry. The present solutions based on mainstream artificial intelligence (AI) struggle, because the problems at hand are too fragmented: the training data is too scarce and the model engineering relies on very specific expert knowledge.
Our Solution, frugal in terms of data and resources, based on a task-agnostic generic device, is able to identify unusual patterns in the analog signals. Our bio-mimicking approach should imitate the ability of human technicians, which assess the state of their machines by the sound. On the long term, our technology could be adapted for a variety of AI applications requiring low energy consumption or full privacy.
The EIC Transition call corresponds exactly to our present needs: accelerate the development and the market readiness of our technology. Moreover, we address explicitly the requirements for Green Digital Devices. By working on the edge, our device reduces the energy and resources required for data transfer and, by imitating the biological neurons, it also reduces the energy required for the computation.
Status
SIGNEDCall topic
HORIZON-EIC-2022-TRANSITIONCHALLENGES-01Update Date
31-07-2023
Geographical location(s)
Structured mapping
Unfold all
/
Fold all
