Explain CLIP, contrastive losses, and retrieval limits

Respond to the machine learning questions below in the setting of multi-modal retrieval over text and video/image:

1. Conceptually, how does a CLIP-style model operate (its architecture, training objective, and how it is used at inference)?
2. Which contrastive learning loss functions are commonly applied to representation learning? Describe several of them and the situations where each is suitable.
3. What are the main limitations of embedding-based retrieval (bi-encoder / vector search)?
4. What other approaches are available (for example, cross-encoders, hybrid sparse+dense, generative retrieval), and what trade-offs does each involve?
5. How would you address or reduce popularity bias in a system that uses embedding-based retrieval?

Overview: This question tests knowledge of multi-modal representation learning and retrieval systems, including CLIP-style joint image–text encoders, families of contrastive losses, the limitations of embedding-based retrieval, alternative retrieval paradigms, and concerns such as popularity bias.