Embedded Files
5.5 year project from Oct 2019 till Mar 2025 for developing a platform for real-time learning at the edge with spiking neural networks.
Intoroduction to our work is given in this article.
5.5 year project from Oct 2019 till Mar 2025 for developing a platform for real-time learning at the edge with spiking neural networks.
Intoroduction to our work is given in this article.
What's the project?
What's the project?
People may imagine, in future, edge devices around us would learn our unconscious behaviours (such as habit, tendency, inclination, and health condition) quietly but in detail. Then, the edge devices will give us “non-intrusive” advice to improve the quality of our life.
People may imagine, in future, edge devices around us would learn our unconscious behaviours (such as habit, tendency, inclination, and health condition) quietly but in detail. Then, the edge devices will give us “non-intrusive” advice to improve the quality of our life.
For this purpose, many of us expect trendy Deep Learning (aka “AI”) would be the most pivotal technology to learn those unconscious behaviours and infer appropriate advice.
For this purpose, many of us expect trendy Deep Learning (aka “AI”) would be the most pivotal technology to learn those unconscious behaviours and infer appropriate advice.
However, the present “AI” requires immense energy for its enormous logic operations. This vast power consumption would never be suitable for wearable edge devices to give users visible or silent support day and night.
However, the present “AI” requires immense energy for its enormous logic operations. This vast power consumption would never be suitable for wearable edge devices to give users visible or silent support day and night.
The internet of things (IoT) may solve the problem. We carry on any power-consuming computations in the cloud computers, i.e., “AI” connected through the high-speed internet. But the edge device, in general, is the source and repository of ample personal information. Therefore, it is not a satisfactory idea to connect everything to the internet.
The internet of things (IoT) may solve the problem. We carry on any power-consuming computations in the cloud computers, i.e., “AI” connected through the high-speed internet. But the edge device, in general, is the source and repository of ample personal information. Therefore, it is not a satisfactory idea to connect everything to the internet.
Then, what is the final solution?
Then, what is the final solution?
The edge device should learn and infer the unconscious behaviours without accessing the internet. For this purpose, we need a platform for on-site ultra-low-power realtime learning/inference. Our idea is to replace the extensive logic calculations in the present machine learning to a spontaneous spatiotemporal pattern formation in the spiking neural network (SNN). The spatiotemporal pattern is called a dynamical attractor. The pattern formation is a kind of phase transition. That is, we replace the gradient method to find the best combination of synaptic weights to the natural providence to avoid the power-consuming logical calculations.
The edge device should learn and infer the unconscious behaviours without accessing the internet. For this purpose, we need a platform for on-site ultra-low-power realtime learning/inference. Our idea is to replace the extensive logic calculations in the present machine learning to a spontaneous spatiotemporal pattern formation in the spiking neural network (SNN). The spatiotemporal pattern is called a dynamical attractor. The pattern formation is a kind of phase transition. That is, we replace the gradient method to find the best combination of synaptic weights to the natural providence to avoid the power-consuming logical calculations.
In this CREST project, we aim to corroborate the scenario above. We try to build our own attractor SNN circuits consisting of our uniquely developed neuromorphic devices. As a demonstration in this project, first, we try to implement the SNN circuits to a kind of personal-authentication device. We keep posting our progress on this website.
In this CREST project, we aim to corroborate the scenario above. We try to build our own attractor SNN circuits consisting of our uniquely developed neuromorphic devices. As a demonstration in this project, first, we try to implement the SNN circuits to a kind of personal-authentication device. We keep posting our progress on this website.
Page updated
Google Sites
Report abuse