Meta-learner that jointly captures multiple data modalities

Develop a meta-learning framework whose meta-learner jointly captures multiple data modalities within a single model, rather than relying on separate modality-specific meta-learners.

Background

The paper notes that meta-training across multiple modalities is typically handled by constructing separate meta-learners, largely due to differences in modality dimensionalities. This compartmentalization limits unified generalization across modalities.

A single meta-learner capable of capturing heterogeneous modalities would strengthen cross-modal transfer and efficiency but remains unresolved.

References

Learning a meta-learner that captures multiple data modalities remains unsolved.

— Towards Incremental Learning in Large Language Models: A Critical Review (2404.18311 - Jovanovic et al., 28 Apr 2024) in Section 2.2 (Meta-Learning)