Compositional Generalization

Published:

Compositional generalization describes a system’s ability to handle new combinations of things it already knows. People do this naturally, such as using familiar words in a sentence they have never heard before or combining learned skills to solve a new problem. Many machine-learning models struggle with this, because they often learn patterns that are tightly tied to the examples they were trained on.

Studying this ability requires careful test design. Simple train–test splits can hide the problem, since the model may see similar combinations during training. Instead, researchers create tests where certain combinations are deliberately held back. Success here is the ability to apply learned rules and meanings in new contexts.

Follow us on Facebook and LinkedIn to keep abreast of our latest news and articles