When children learn a new object, such as scissors or a bowl, they may turn it around, ask what it is, and then be able to recognize it again in various settings and conditions without confusing it with other objects learned at about the same time.
It is a kind of magic, as researchers in neuromorphic computing noted in a recent paper, “Interactive continual learning for robots: a neuromorphic approach,” presented by Intel Labs in collaboration with the Italian Institute of Technology and Technical University of Munch. The researchers are trying to apply the same childlike approach to interactive continual learning for smart robots, especially those used as robotic assistants that interact with other robots and people in healthcare, elderly care or logistics.
The group developed models and was able to successfully demonstrate continual interactive learning on Intel’s Loihi neuromorphic research chip with far less energy usage (up to 175 times lower energy) and with similar or better speed and accuracy than on a conventional CPU.
Intel said their work is important in improving capabilities for manufacturing or assistive robots used in neuromorphic computing as they adapt to unforeseen conditions and work in an agile manner alongside humans.
In one simulation, a robot sensed objects by moving its eyes made of event-based cameras, also known as Dynamic Vision Sensors, to generate microsaccadic events used to drive a spiking neural network on a Loihi chip. If the object was new, its SNN representation was learned or updated. If the object was known, it was recognized by the network and that feedback was given to the user.
The researchers noted the need to test their approach on actual robots and many more objects than were used in the project.
The research won “Best Paper” in July at the International Conference on Neuromorphic Systems hosted by Oak Ridge National Laboratory.
One of the principal authors, Yuliya Sandamirskaya, leads the applications research group of the Intel Labs Neuromorphic Computing Lab. She is based in Munich Germany.
Por: Matt Hamblen