What is adaptive computation time (ACT) in deep vision models?

ACT allows models to dynamically decide how many computation steps are needed per input.

Adaptive Computation Time enables neural networks to allocate more computation to complex inputs and less to simpler ones. A halting unit predicts when sufficient computation has been performed. This is particularly useful in vision tasks where some images are easy (clear objects) and others are hard (occlusion, noise). It improves efficiency without sacrificing accuracy.