A projection layer is the intermediate layer in a Word2Vec network whose rows become learned word vectors. Training uses these vectors to predict context or center words from raw text; after training, the layer is discarded and the dense embedding matrix remains. CBOW sums context-word embeddings for efficient prediction, while Skip-gram predicts surrounding words independently and often captures stronger relationships for infrequent terms. section 2.1.2
Figure 2.1 CBOW vs skip-gram architectures.

Compare how CBOW and Skip-gram route words through the projection layer before producing predictions. section 2.1.2
