University of Toronto Algorithm Learns From Instructions, Outperforms Standard Training

This article was prepared from the source material provided, which includes the original article text and details from the study published in IEEE Transactions on Neural Networks and Learning Systems.

Engineers Parham Aarabi and Wenzhi Guo at the University of Toronto have developed a machine learning algorithm that learns from human instructions rather than labeled datasets, according to a study published in IEEE Transactions on Neural Networks and Learning Systems. The heuristically trained neural network outperformed conventional machine learning algorithms by 160 percent and exceeded its own training reliability by 9 percent when identifying hair in photographs.

Most AI systems learn by exposure to existing sets of examples. The Toronto team's approach, called "heuristic training," instead relies on direct instructions—for example, telling the algorithm that "sea water is likely to be in shades of blue" rather than showing it hundreds of photos of water. The researchers applied this method to the task of identifying hair in photographs.

"Our algorithm learned to correctly classify difficult, borderline cases — distinguishing the texture of hair versus the texture of the background," Aarabi said. "What we saw was like a teacher instructing a child, and the child learning beyond what the teacher taught her initially."

The algorithm's ability to improve beyond its initial training suggests potential for classifying previously unknown or unclassified data. Guo said the team is "keen to apply our method to other fields and a range of applications, from medicine to transportation," including identifying previously unseen cancerous tissues during medical analysis or objects surrounding an autonomous vehicle.