Early fusion concatenates or otherwise combines raw or lightly processed features from each modality before they enter the main model pipeline, producing a single joint input representation that a downstream network learns from jointly. This contrasts with late fusion (option C), where separate unimodal models process each modality independently and their outputs — logits, embeddings, or decisions — are combined only at the end, and with intermediate/hybrid fusion, where combination happens at one or more intermediate feature layers.

Early fusion's advantage is that it allows the model to learn cross-modal correlations from the earliest layers, potentially capturing low-level interactions that later fusion stages would miss. Its disadvantage is sensitivity to modality-specific noise, differing sampling rates, and missing modalities: if one input stream is corrupted or absent, the joint representation degrades more severely than in late fusion, where the surviving modality's model can still function independently.

Option B describes unimodal reduction, not fusion at all, and option D confuses a data-processing strategy with a project-management timeline — an easy distractor to eliminate. Exam questions frequently test the ability to distinguish early, late, and hybrid fusion by identifying *where* in the pipeline combination occurs, so anchor your answer to pipeline stage rather than performance characteristics.