Early Fusion
A strategy in multimodal or mixed-feature modeling in which different features are directly concatenated at the input layer before being fed into the model. It is simple and efficient but may make it difficult for the model to learn interactions due to differences in feature scales.
Timeline (last 90 days)
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.
Early fusion strategies can capture cross-modal fine-grained interactions, but they impose higher demands on input length and computational resources, and practical deployment needs to pay attention to video frame sampling strategies and context window limits.
Early fusion means mapping tokens of different modalities into the same representation space at the model input stage and letting them interact within the same Transformer.
All Facts (3)
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.
50%UnverifiedEarly fusion strategies can capture cross-modal fine-grained interactions, but they impose higher demands on input length and computational resources, and practical deployment needs to pay attention to video frame sampling strategies and context window limits.
50%UnverifiedEarly fusion means mapping tokens of different modalities into the same representation space at the model input stage and letting them interact within the same Transformer.
50%