<ul><li>SEAL is a new framework designed for data-incremental learning, where data samples arrive sequentially and are not stored for future access.</li><li>SEAL dynamically adapts the model structure by expanding it only when necessary, based on a capacity estimation metric, to balance plasticity and stability in incremental learning.</li><li>The framework uses Neural Architecture Search (NAS) to search for both the optimal architecture and expansion policy, preserving stability through cross-distillation training.</li><li>Experiments show that SEAL effectively reduces forgetting, enhances accuracy, and maintains a lower model size compared to existing methods, promising efficient adaptive learning in incremental scenarios.</li></ul>

SEAL: Searching Expandable Architectures for Incremental Learning

Discover more