Current AI systems are quite high-maintenance. They require high computing power, extensive memory and a networked environment. Researchers at the University of Waterloo want to train them to make do with fewer resources and are attempting to do so using processes inspired by biological evolution: They are gradually reducing memory space and computing power, and they are depriving the system of an Internet connection – with amazing success: The AI software required for object recognition shrank over the course of time to two-hundredths of its original size.
The application areas for such compact AI systems are diverse. They may prove useful in mobile phones, as these often rely on Internet connections with widely varying bandwidths. The stripped-down AI systems could also provide valuable services for industrial robots or in areas where data protection is critical. With its Go-playing AI AlphaGo Zero , Google has recently demonstrated just how powerful AI systems based on deep learning could become.