Researchers from the Centrum Wiskunde & Informatica have developed a computer algorithm that can theoretically recognize heart defects, speech and hand gestures using pulsed neural networks up to a thousand times more efficiently than traditional AI techniques.
CWI researcher Bojian Yin and professor Sander Bohté made their discovery together with Federico Corradi of the Interuniversity Micro-Electronics Center Foundation in Eindhoven. They achieved a ‘mathematical breakthrough’ in computing with so-called ‘pulsed neural networks’ or spiking neural networks.
With this discovery, special neuromorphic chips developed for this kind of artificial intelligence can recognize speech, gestures and electrocardiograms for heart defects “a factor of twenty to a thousand” more energy-efficient than traditional AI applications. They have published an article about this in the journal Nature Machine Intelligence.
The three researchers came up with a mathematical solution that makes neural networks more similar to those of the human brain. Normal neural networks work with continuous signals. Pulsed neural networks calculate with pulses, just like the human brain. The disadvantage, the researchers explain, is that the signals are discontinuous and therefore mathematically ‘more difficult to handle’.
The researchers developed a computer algorithm that can handle these pulsed neural networks. Several benchmark tests conducted by the researchers showed that the algorithm performs at least as well as traditional deep neural networks, but is much more energy-efficient. In theory, a factor of a hundred to a thousand, according to the researchers.
In collaboration with imec, the researchers developed a neuromorphic computer chip based on their algorithm with 336 pulsed neurons: the Brain chip. Because the chip is much more energy efficient than the current generation of AI applications for speech, gestures and heart rhythm recognition, it can be used, for example, to detect heart defects by implanting them. In theory, it would run on a single battery for a year, the researchers say.
μBrain chip from IMEC. Image: imec