A math idea that may dramatically reduce the dataset size needed to train AI systems | Machine learning & AI - Tech Xplore

A pair of statisticians at the University of Waterloo has proposed a math process idea that might allow for teaching AI systems without the need for a large dataset, as Tech Xplore reports.

Photo: CC0 Public Domain

Ilia Sucholutsky and Matthias Schonlau have written a paper describing their idea and published it on the arXiv preprint server.

Artificial intelligence (AI) applications have been the subject of much research lately, with the development of deep learning networks, researchers in a wide range of fields began finding uses for it, including creating deepfake videos, board game applications and medical diagnostics.

Deep learning networks require large datasets in order to detect patterns revealing how to perform a given task, such as picking a certain face out of a crowd. In this new effort, the researchers wondered if there might be a way to reduce the size of the dataset. They noted that children only need to see a couple of pictures of an animal to recognize other examples. Being statisticians, they wondered if there might be a way to use mathematics to solve the problem.

The researchers built on recent work by a team at MIT.

 Read more...

Additional resources 

'Less Than One'-Shot Learning: Learning N Classes From M < N Samples, arXiv:2009.08449 [cs.LG]

https://arxiv.org/abs/2009.08449

Source: Tech Xplore