Two-Way Is Better Than One: Bidirectional Alignment with Cycle Consistency for Exemplar-Free Class-Incremental Learning
Researchers have developed a new method called BiCyc for exemplar-free class-incremental learning, which aims to prevent models from forgetting previously learned information when acquiring new skills. Existing projection-based methods can introduce biases by distorting feature geometry or only aligning old classes locally. BiCyc addresses this by using a bidirectional projection approach with a cycle-consistency objective, jointly optimizing two maps to allow transport and representation to co-evolve. This method demonstrably reduces forgetting and improves accuracy on standard benchmarks. AI
IMPACT Improves model ability to learn new tasks without forgetting old ones, crucial for long-term AI development.