🔀Late Fusion

NEW

A multimodal strategy that combines outputs or decisions from separate models or modalities at the final decision stage.

DESCRIPTION

Late fusion processes each modality independently and merges their predictions via techniques like averaging, weighted voting, stacking, or learned gating. It offers modularity and robustness to heterogeneous modalities, simplifying integration when feature spaces differ. However, it may miss cross-modal interactions captured by early or intermediate fusion. Late fusion is common in audio-visual tasks, multi-sensor systems, and ensembles.

ALTERNATIVE DEFINITIONS

Wikipedia

Decision-level fusion (late fusion) combines the decisions of multiple classifiers.

Defines late fusion specifically as decision-level combination.

Source

Wikipedia

In late fusion, modalities are processed independently and their outputs are combined at the decision level.

Highlights independent processing and decision-level merging.

Source

RELATED TERMS

Multimodal Learning Ensemble Learning Decision-level Fusion Model Averaging Stacked Generalization Voting Classifier Feature Fusion Sensor Fusion

COMPETING TERMS

Early Fusion Intermediate Fusion

KNOWLEDGE AREAS

AI Data Science

TARGET AUDIENCE

AI Engineer Data Scientist

EXTERNAL SOURCES

• Wikipedia - Data fusion
• CVPR Open Access - Survey on multimodal fusion strategies.
• scikit-learn - Ensemble methods for combining model outputs.

TOP QUESTIONS

When is Late Fusion preferable to Early Fusion?

When modalities have incompatible feature spaces or different preprocessing, late fusion allows modular training and straightforward integration.

Category: Opinion

How do you implement Late Fusion?

Train separate models per modality and combine predictions via averaging, weighted voting, stacking, or learned gates.

Category: Example

What is a common mistake with Late Fusion?

Ignoring calibration across modalities leads to biased combinations; apply calibration or weight predictions appropriately.

Category: Mistakes

What risks exist with Late Fusion?

It can miss cross-modal interactions and synergies that feature-level fusion captures, limiting performance.

Category: Risks

What’s a hot topic around Late Fusion?

Hybrid fusion architectures combining early, intermediate, and late fusion with attention or gating mechanisms.

Category: Topic

How do you evaluate Late Fusion systems?

Compare modality-specific models versus fused outputs, measure calibration, and analyze contributions via ablations.

Category: Testing

What’s a striking analogy for Late Fusion?

It’s like a jury where each expert votes based on their specialty, and the final verdict merges their decisions.

Category: Example

Does Late Fusion require more data?

Not necessarily; it often reuses modality-specific datasets but may need validation data for fusion weights.

Category: Opinion

Can Late Fusion improve robustness?

Yes, diverse modalities and models reduce single-source failure, enhancing robustness and reliability.

Category: Relevancy