CONCEPT Cited by 1 source

Early Fusion (Multimodal)

Early fusion is the multimodal-ML strategy of concatenating raw (or minimally-encoded) modality inputs into a single feature vector before training, so one joint model learns directly over the combined input. It sits at the earliest possible fusion point in the pipeline — upstream of any modality-specific representation learning.

When it survives production

"Small, tightly controlled cohorts with consistent modality availability" (Source: sources/2026-04-22-databricks-multimodal-data-integration-production-architectures-for-healthcare-ai).

The requirements are strong because early fusion gives up two things other strategies keep:

• No graceful degradation on missing modalities. Zero-padding a concatenated input lets the model silently learn that zeros are informative; at deployment, a genuinely missing modality is indistinguishable from a present-but-zero one.
• Dimensionality imbalance swamps the learner. High- dimensional inputs (genomics variants) drown lower-dimensional ones (structured EHR fields) unless per-modality scaling + regularisation is engineered carefully.

Failure mode called out in ingested sources

"Scales poorly with high-dimensional genomics and large feature sets." (Source: sources/2026-04-22-databricks-multimodal-data-integration-production-architectures-for-healthcare-ai.) Genomics profiling can contribute ≥10⁶-dim feature vectors; EHR features sit in the 10²–10³ range. Direct concatenation hands the model a 1000:1 dimensionality gradient to fight through.

Contrast with other fusion strategies

• concepts/intermediate-fusion — encode each modality separately, then merge hidden representations. Buys dimensionality-balance at the cost of more training discipline.
• concepts/late-fusion — train per-modality models, combine predictions. Buys graceful degradation on missing modalities at the cost of ignoring cross-modal interactions until the very end.
• concepts/attention-based-fusion — dynamic weighting across modalities and time. Buys temporal / interaction modelling at the cost of validation complexity.

See patterns/fusion-strategy-selection-by-deployment-reality for the decision-framework shape Databricks' post frames the fusion choice inside.

Seen in

• sources/2026-04-22-databricks-multimodal-data-integration-production-architectures-for-healthcare-ai — Databricks names early fusion as the first of four fusion strategies and immediately scopes it to small / tightly- controlled cohorts with consistent modality availability; warns it "scales poorly with high-dimensional genomics and large feature sets". The post's decision framework — "match fusion to your deployment reality" — uses early fusion as the canonical example of a strategy that works in research cohorts and breaks under production sparsity.