In the context of multimodal machine learning, what does 'data fusion' refer to?

Data fusion is the process of combining information from multiple modalities into a single, unified representation that a downstream model can act on. As covered in the early-fusion and late-fusion questions elsewhere in this set, fusion can occur at different pipeline stages — raw/feature-level (early), intermediate representation level (hybrid), or decision level (late) — but in every case the defining operation is combination, not separation.

Option A describes the inverse operation and does not correspond to any standard multimodal technique under the name "fusion." Option C describes missing-data handling or imputation, a data-quality concern that is often addressed *before* fusion (a model needs some representation for each modality, even an imputed or masked one, before combining them) but is not fusion itself. Option D describes evaluation or quality assessment, a distinct concern from the mechanical act of combining modalities into one representation.

Fusion technique choice has real architectural consequences: early fusion assumes tight temporal/spatial correspondence between modalities and is sensitive to missing streams; late fusion is more robust to missing or noisy modalities since each unimodal branch can still contribute independently; hybrid/intermediate fusion, common in modern transformer-based multimodal architectures via cross-attention, aims to capture the benefits of both while mitigating each one's weaknesses.