Explain multimodal fusion techniques — early fusion versus late fusion.

Fusion is how a model combines information from multiple modalities. Early (feature-level) fusion merges the modalities near the input — concatenating or jointly attending over their embeddings — so the model learns rich cross-modal interactions from the start; modern VLMs that feed visual tokens into an LLM are essentially early/intermediate fusion. It's powerful for tasks needing tight interplay (visual question answering) but needs aligned, synchronised data and is heavier to train. Late (decision-level) fusion processes each modality separately with its own model and combines only the final outputs — averaging scores or a small combiner. It's simpler, modular, robust to a missing modality, and lets you reuse unimodal models, but it can't capture fine cross-modal interactions. Many systems use intermediate/hybrid fusion to balance the two.