Dual-Encoder Architecture
A neural network architecture design consisting of two independent encoders that each process different modalities or input types, matching them by similarity measures within a shared vector space. The architecture supports pre-computation and caching of embeddings to enable efficient retrieval.
Timeline (last 90 days)
The dual-tower architecture can pre-compute and cache the embeddings of one side, so at query time only the other side needs to be encoded followed by vector retrieval, resulting in extremely low computation; this is the architectural root of CLM's speed advantage, with the trade-off being that the two-sided encoders cannot perform fine-grained cross-attention comparison.
The Two-Tower architecture uses two separate encoders to map queries and documents into the same low-dimensional vector space, measuring relevance through vector similarity.
All Facts (2)
The dual-tower architecture can pre-compute and cache the embeddings of one side, so at query time only the other side needs to be encoded followed by vector retrieval, resulting in extremely low computation; this is the architectural root of CLM's speed advantage, with the trade-off being that the two-sided encoders cannot perform fine-grained cross-attention comparison.
50%UnverifiedThe Two-Tower architecture uses two separate encoders to map queries and documents into the same low-dimensional vector space, measuring relevance through vector similarity.
50%