Transformers and LLMs

Thin Modality Towers, explained

Thin modality towers are lightweight modality-specific encoders that project inputs such as images or audio into a shared model space.
Back to Glossary

Part of the Semantic Notion technical glossary.

Short Definition

Thin modality towers are lightweight modality-specific encoders that project inputs such as images or audio into a shared model space.

Intuition

The modality adapter stays relatively simple, while the large shared model does more of the representation and reasoning work.

Technical Definition

A thin modality tower uses comparatively small preprocessing or projection modules to map non-text inputs into embeddings consumed by a shared decoder or backbone.

Example

Inkling uses relatively simple image and audio towers before projecting modalities into a shared hidden space.

Common Misunderstandings

Thin modality towers do not mean multimodal understanding is simple.
They are an architectural tradeoff, not a universal best practice.

Start here

Need the broader concept map?

Return to the glossary index to move from this definition into adjacent AI, ML, mathematics, and computer science terms.