Optimization as a model for few-shot learning by S Ravi and H Larochelle at ICLR 2017

Interesting new research on how to effect few-shot learning.

AI researchers are making progress in the area of one-shot or few-shot learning. The direction of travel has been to ally a narrow-focus learning algorithm with an algorithm that learns at a less granular level across a number of areas or classes. The second algorithm aims to identify similarities within classes so as to weed out whole classes that are clearly wrong. Such an algorithm might, for example, put all animals in a class. When it sees an elephant, it quickly discards alternative classes such as trees or bodies of water. Stage 1 of AlphaGo Fan’s approach cuts the search focus to a manageable number of promising solutions. Artificial few-shot learning similarly weeds out clearly incorrect solutions in large clumps. The authors propose their meta-learning technique as a viable alternative to metric learning techniques.

Link to paper: https://openreview.net/pdf?id=rJY0-Kcll

You may also like to browse other AI papers: https://www.thesentientrobot.com/category/ai/ai-academic-papers/