Representation
How a model re-encodes something into its own internal language of numbers.
Think of it like
Sheet music for a song — not the sound itself, but a form the performer can actually work from.
Example
A photo of a cat becomes a representation — a stack of numbers — before the model can classify or caption it.
How it actually works
Deep learning’s whole trick is learning good representations: turning raw input into vectors where the useful structure is easy to read off. Better representations make downstream tasks easier. Much of what pretraining buys you is reusable representations.
For product teams
The quality of a model’s internal representations sets the ceiling on everything it can do downstream.
For engineers
A learned feature vector encoding input; the output of intermediate layers used by downstream heads.
Related
- Vector — Stored as a vector.
- Embeddings — The learned kind is an embedding.
- Deep learning — Learning good ones is deep learning.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome