Late Fusion
A strategy in multimodal or mixed-feature modeling in which each type of feature is encoded independently and interactions occur at a higher level. It better preserves the characteristics of each modality but incurs greater computational overhead.
Timeline (last 90 days)
This study proposes using a supervised large language model (LLM) for late fusion, aggregating the outputs of heterogeneous echocardiography AI models in language space, eliminating the need for format conversion and feature engineering for each individual model.
Late fusion means each sub-model first completes its prediction independently, and then the results are merged at the output level; early fusion instead concatenates multi-source information at the feature level for joint training.
All Facts (2)
This study proposes using a supervised large language model (LLM) for late fusion, aggregating the outputs of heterogeneous echocardiography AI models in language space, eliminating the need for format conversion and feature engineering for each individual model.
50%UnverifiedLate fusion means each sub-model first completes its prediction independently, and then the results are merged at the output level; early fusion instead concatenates multi-source information at the feature level for joint training.
50%