<ul><li>Out-of-distribution generalisation is crucial for human and animal intelligence.</li><li>To achieve OOD through composition, an intelligent system must identify task-invariant input features and composition methods.</li><li>Testing on an OOD setup is not sufficient; confirming that features are compositional is also essential.</li><li>Exploration of tasks shows that some neural networks struggle with OOD, while novel architectures with appropriate biases can be successful.</li></ul>

Out-of-distribution generalisation is hard: evidence from ARC-like tasks

Discover more