Positional Encoding
How a transformer, which has no inherent order, learns where tokens are.
Overview
Positional encoding injects information about a token's position, because attention itself is permutation-invariant and would otherwise see a bag of words. It solves the problem that order carries most of the meaning in language and none of it is present in the mechanism. It exists as one of the few places where the original architecture has been substantially revised, since how position is represented decides how well a model extrapolates to sequences longer than it was trained on.
Technology Graph
How Positional Encoding connects to the rest of the stack.
Classification
- Transformer architecture Algorithms Positional Encoding part of Transformer High ✓ source
Tools
Software that implements or integrates with Positional Encoding.