Energy-efficient AI detects heart defects – CWI Amsterdam

CWI researchers Bojian Yin and Sander Bohté, as well as their colleague Federico Corradi from Stichting Interuniversitair Micro-Elektronica Centrum (IMEC) in Eindhoven, achieved a mathematical breakthrough in the computation of spike neural networks.

Thanks to this breakthrough, special chips adapted to this artificial intelligence (AI) can recognize speech, gestures and electrocardiograms (ECGs) a factor of twenty to a thousand more efficiently than traditional AI techniques. Such chips are on the eve of practical and everyday applications.

The research results have been published in the scientific journal Nature Machine Intelligence October 14, 2021.

IMEC brain-chip (Credit: IMEC)

Energy saving

Over the past decade, AI has gained more and more everyday applications, especially for recognizing images and spoken words. This is done with deep neural networks, which are very simplified mimics of how the human brain processes information. For mobile apps, however, running current AI models often costs too much energy. Developing low-power AI has therefore become increasingly important.

One way to make AI applications more energy efficient is to make neural networks more similar to those in the human brain. Classical neural networks use continuous signals that are easy to manipulate mathematically. Spiked neural networks calculate with pulses, which is much more like what’s going on in the brain and takes less energy, but has the downside that the signals are discontinuous and more difficult to handle mathematically. However, Bohté and his two co-authors found a mathematical solution to this problem.

“We tested our computer algorithm on three benchmarks,” explains Bohté. “These benchmarks consist of a series of tests of ten gestures, a series of words and a continuous ECG signal. Our algorithm is at least as efficient but much more energy efficient than traditional deep neural networks. In theory, we gain a factor of one hundred to one thousand. “

Computer brain

To use algorithms like Bohté’s in everyday applications, special neuromorphic computer chips are needed. The architecture of these chips is more like the biological architecture of the human brain than that of traditional computer chips. Bohté: “Based on our algorithms, our research partner IMEC has manufactured a special neuromorphic chip with 336 spiked neurons: the μBrain chip. If we run our algorithm on this special chip, we gain a factor of 20 in power consumption. theoretical energy gain, the practical energy gain is always lower because of the conversion of digital signals to analog and vice versa, and because of the reading of the data. But a 20-fold energy gain is still a lot. To detect heart defects, that means you can implant an ECG recording chip and it will run for a year on a single battery. “

In the coming years, neuromorphic chips will contain more and more spiked neurons, which will further expand the possibilities of applying artificial intelligence in portable chips. For example, at the end of September, the American chipmaker Intel produced the neuromorphic chip Loihi 2, which already contains a million spiked neurons.

The research project of Bohté and his colleagues is taking place within the framework of the NWO Perspective “Efficient Deep Learning” program.

More information