liveBook · Manning

Permutation invariance

As explained in Applied Reinforcement Learning

The property that reordering inputs changes the output only by the corresponding reordering, without changing relational results.

Permutation invariance means reordering inputs reorders outputs in the same way, without changing their relationships. Attention treats inputs as a set rather than a sequence, so zone identity and spatial arrangement may need to be added explicitly. section 9.6

Swapping zones

Swap zones 2 and 5 in the input: the outputs swap correspondingly. To represent each zone’s location, add a unique positional signal to its representation before it enters the Transformer. section 9.6

How Applied Reinforcement Learning covers it

How deep it goes
Explains set-like behavior and addresses it with positional signals and learnable relative-position patterns.
Examples
Shows zones 2 and 5 swapping, with their outputs swapped correspondingly.
What you can do after
Understand why attention may need positional signals when zone identity and spatial arrangement matter.
How it connects
Connects attention and Transformers to positional encoding and relative position.

Also explained in

Compare how 4 books explain permutation invariance →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Applied Reinforcement Learning. The section links point to the parts of the book it draws on.