Fig. 2: Fusion strategies using deep learning.

Model architecture for different fusion strategies. Early fusion (left figure) concatenates original or extracted features at the input level. Joint fusion (middle figure) also joins features at the input level, but the loss is propagated back to the feature extracting model. Late fusion (right figure) aggregates predictions at the decision level.