<ul><li>Certifying the IID generalisation ability of deep networks is crucial for trusting AI in high-stakes applications like medicine and security.</li><li>Obtaining non-vacuous guarantees for generalisation bounds with contemporary large models on small-scale data is challenging.</li><li>A novel connection is made between learning methods based on model fusion and generalisation certificates, showing existing strategies can provide non-trivial generalisation guarantees.</li><li>Learning with as low as 100 examples using models like VIT-B and mistral-7B can now have non-trivial generalisation guarantees, enhancing trustworthiness of AI systems.</li></ul>

Model Merging is Secretly Certifiable: Non-Vacuous Generalisation Bounds for Low-Shot Learning

Discover more