Compressed neural networks lose the ability to adapt on their own even when they perform well with guidance.

Real-world AI models must be small and able to adjust to new situations. To make them small, engineers compress them. To help them adjust on the fly without extra labels, they use test-time adaptation. But we do not understand how these two steps affect each other. Researchers tested compressed models to see how well they handle unexpected data changes. When models adapt using labeled examples, compressed models still get high accuracy. However, adapting on the fly without labels fails as compression grows heavier. Squishing the model shrinks the diversity of what it can represent and locks it into rigid structures. Because different compression methods cause different amounts of damage, future compression designs must protect a model's ability to adapt.

Why it matters

If you shrink a neural network to run faster on a phone or edge device, it might fail when conditions change unexpectedly. This means AI that needs to adapt in real time, like in self-driving cars or smart assistants, could become less reliable when optimized for speed.

Questions answered