Bookworm continual learning: beyond zero-shot learning and continual learning
This addresses the challenge of flexible and continual learning for AI systems that need to handle evolving data with new classes, though it appears incremental as it builds on existing CL and ZSL methods.
The paper tackles the problem of learning from both seen and unseen classes over time by introducing bookworm continual learning (BCL), which generalizes continual and zero-shot learning, and proposes the bidirectional imagination (BImag) framework with variants to improve feature generation, achieving competitive results on benchmarks.
We propose bookworm continual learning(BCL), a flexible setting where unseen classes can be inferred via a semantic model, and the visual model can be updated continually. Thus BCL generalizes both continual learning (CL) and zero-shot learning (ZSL). We also propose the bidirectional imagination (BImag) framework to address BCL where features of both past and future classes are generated. We observe that conditioning the feature generator on attributes can actually harm the continual learning ability, and propose two variants (joint class-attribute conditioning and asymmetric generation) to alleviate this problem.