The ability to extract predictions or representations from intermediate layers of a neural network rather than only from the final output, allowing trade-offs between speed and accuracy.