Researchers at Eindhoven University of Technology are paving the way for efficient and dedicated AI semiconductors that will enable on-chip training of AI models.
Researchers at Eindhoven University of Technology (TU/e) have developed a neuromorphic device that enables on-chip training and eliminates the need to transfer trained models to the chip. This could open the way to efficient and dedicated AI chips. Details can be found in “Science Advances”.
ECRAM as a basis
“Neural networks can help solve complex problems with large amounts of data. But the larger the networks become, the higher the energy costs and hardware limitations. But there is a promising hardware-based alternative – neuromorphic chips,” emphasizes TU/e scientist Yoeri van de Burgt.
The experts created a two-layer neural network based on ECRAM components made of organic materials and tested the hardware with a further development of the widely used backpropagation training algorithm.
“The conventional algorithm is often used to improve the accuracy of neural networks, but is not compatible with our hardware, so we developed our own version,” says van de Burgt’s colleague Tim Stevens.
Two-layer network
“We have shown that this works for a small two-layer network. Next, we want to involve industry and other large research labs so that we can build much larger networks of hardware devices and test them with real data problems,” said van de Burgt.
This would make it possible to demonstrate that these systems are very efficient in training and operating useful neural networks and AI systems. “We would like to apply this technology in several practical cases. My dream is that such technologies will become the norm in AI applications in the future,” says the researcher.
Source: www.com-magazin.de
