New Data Learning Methods for AI

IoT technology allows us to easily and continually obtain large amounts of diverse data, and AI is gaining attention as a way to use that big data.

Machine learning mainly deals with single-label classification problems so that data and corresponding label information are in a one-to-one relationship. Data and label information seldom have a one-to-one relationship in the real world. Multi-label classification problems deal with data with a one-to-many relationship between data and label information. To efficiently learn from big data obtained continually, the ability to learn over time without destroying things that were previously learned is also required.

A research group from the Osaka Metropolitan University Graduate School of Informatics developed a new method that combines classification performance for data with multiple labels with the ability to learn with that data continually. Numerical experiments on real-world multi-label datasets indicate that the proposed method outperforms conventional methods.

The new algorithm makes it easy to devise an evolved version that can integrate with other algorithms. By learning the data and the label information corresponding to the data separately and continually, they achieve high classification performance and continual learning capability.

The research results were published in IEEE Transactions on Pattern Analysis and Machine Intelligence on December 19, 2022.