Researchers train AI to recognise objects more efficiently
Computing & Tech
Advancements in AI training could open its benefits up to sectors, companies and individuals who don’t have access to large datasets
Spotted: A team at the University of Waterloo in Ontario has developed a new method for training AI to identify categories, based on degrees of similarity.
The Waterloo team has developed what they call “soft labels”. Unlike “hard labels”, which label a data point as belonging to one specific class, soft labels identify the level of analogy between that data point and multiple classes. For example, an AI only trained on dogs and cats would be able to identify a third class of animal, such as kangaroos, by describing it as “60 per cent like a dog and 40 per cent like a cat”.
“Translation? Tell an AI a kangaroo is some fraction cat and some fraction dog (both of which it’s seen and knows well) and it’ll be able to identify a kangaroo without ever having seen one,” said the team.
If the soft labels are detailed enough, it could be possible to teach an AI to identify a large number of categories based on just a few training examples.
The paper’s authors use a simple algorithm called K-Nearest Keighbors (kNN) to differentiate between categories. The team chose specific features to represent each category, such as weight and size for animals. This allows the soft label features to be plotted as data points on a graph. With one feature “wight” on the X-axis and the other “size” on the Y-axis, data points that are similar to each other are clustered close by. Based on the distribution of each prototype, identification can be provided.
Written By: Katrina Lane
Explore more: Computing & Tech Innovations | Science Innovations
5th November 2020
Email: isucholu@uwaterloo.ca
Website: arxiv.org