computer-vision

Contrastive image-text embedding (the CLIP-style mechanism)

taught

Training paired image and text encoders so a matching image-caption pair's embeddings land close together and every mismatched pair in the batch lands far apart, producing one shared embedding space for both modalities.

Taught in

Contrastive image-text embedding
Computer Vision
Embeddings and vision-language models -- current practice
Computer Vision

Requires

transformer-architecture recommended
deep-learning ch.18 derives the Transformer encoder these image and text towers are typically built from, in full; this course's own quizzes only require the working-depth version taught locally.
self-attention recommended
deep-learning ch.17 derives self-attention in full; this course's own quizzes only require the working-depth version taught locally.

Required by

Multimodal embeddings and multimodal RAG recommended

Local knowledge map

Click to expand

Solid arrows are required prerequisites; dashed are recommended or helpful. Lower rows are more foundational. Click any box to jump to that concept.