Late Fusion
Processing each modality separately, then combining predictions or embeddings - learn each independently, then fuse.
Think of it like
Each person reads the summary alone, then they discuss and agree on the interpretation.
Example
Image classifier outputs "dog (0.9)" and text classifier outputs "dog (0.8)" -> ensemble: average to 0.85.
How it actually works
Pro: modalities processed independently (easier debugging, transfer learning), robust to missing modalities (can skip one branch). Con: misses early cross-modal cues, duplicates computation. Typically: separate encoders per modality, combine embeddings or predictions. Fusion can be: averaging, attention-based, gated, or learned weights.
For product teams
Flexible, handles missing modalities, easier to debug.
For engineers
Separate encoders, then fusion layer (attention, gating, or simple averaging). Metric: depends on task.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome