Learns to Remember: Topology-Guided Transfer Boosts Cross-Domain Action

Domain shift is a tough problem in cross-domain action recognition: models that learn one visual setting struggle when cameras, lighting, rooms, or people change. Researchers at Tokyo Metropolitan University propose a topology-guided transfer learning framework aimed at this gap. Published in the International Journal of Machine Learning and Cybernetics, the work by Yuning Shi and Naoyuki Kubota replaces generic fine-tuning or distribution alignment with a graph-based “memory” of the source domain. Using the growing neural gas (GNG) algorithm, they distill the geometric structure of source features into an unsupervised network of prototype nodes and edges. During adaptation, that frozen topology selectively informs learning, outperforming baselines across five backbone architectures and two transfer tasks. The approach is inspired by how humans reuse only relevant prior knowledge.