Intel and the Italian Institute of Technology, study on neural networks and AI

A crucial step in improving the service or production capabilities of future robots

Intel Labs, in collaboration with the Italian Institute of Technology and the Technical University of Munich, presented a new approach to object learning based on neural networks. It is particularly targeting future applications such as robotic assistants interacting with unconstrained environments, including logistics, healthcare or elderly care. This research is a crucial step in improving the service or manufacturing capabilities of future robots. Neuromorphic computation is employed through new interactive online object learning methods to allow robots, after their release, to discover new objects.

Using these new models, Intel and other researchers successfully demonstrated interactive learning on the Loihi neuromorphic chip, achieving up to 175 times lower power consumption when learning new object instances, with similar or better speed and accuracy than conventional methods. run on CPU. To achieve this, the researchers implemented a spiking neural network architecture on Loihi, which made it possible to localize the learning of the object in a single layer of plastic synapses and collected the different perspectives of the object involving other neurons on request. . In this way the learning process took place autonomously during the interaction with the user.

This research was published in the article “Interactive continual learning for robots: a neuromorphic approach,” which was named “Best Paper” at this year’s International Conference on Neuromorphic Systems (ICONS), held by Oak Ridge National Laboratory. “When a human being knows a new object, he takes a look, turns it, asks what it is and then is able to instantly recognize it again in all kinds of contexts and conditions in which he finds it,” he said. Yulia Sandamirskaya, head of robotics research in Intel’s Neuromorphic Computing lab and senior author of the paper. “Our goal is to introduce robotic-like capabilities of the future that will act in interactive environments, enabling them to adapt to situations they are unable to predict and to work more naturally with humans. Our results with Loihi reinforce the value of Neuromorphic Computing for the future of robotics. “