<ul><li>Weight imprinting is a universal and efficient method for adaptation to new downstream tasks.</li><li>A framework for weight imprinting is proposed, consisting of generation, normalization, and aggregation.</li><li>The benefits of using multiple proxies for representing novel data and the importance of proper normalization are highlighted.</li><li>A novel variant of imprinting that outperforms previous work is introduced, with an increase of up to 4% in challenging scenarios.</li></ul>

Robust Weight Imprinting: Insights from Neural Collapse and Proxy-Based Aggregation

Discover more