Browse Teach Curate

Balance-aware Sequence Sampling in Multimodal Learning

A training-order strategy for multimodal classification that mitigates modality imbalance by scoring each sample's cross-modal balance and presenting samples from balanced to imbalanced. Students learn how a balance score (prediction similarity plus training loss) drives either a fixed curriculum pacing function or an epoch-updated probabilistic sampler, and why ordering training data can rebalance a multimodal model without extra modules.

Data Text

Balance-aware Sequence Sampling Makes Multi-modal Learning Better Zhi-Hao Guan

This paper introduces Balance-aware Sequence Sampling (BSS), a training-order strategy for multimodal classification that combats modality imbalance - the tendency of a jointly trained multimodal net…