Change on new inputs
| gender | tense |
|---|
| edits applied | | |
| wanted | | |
- Mixing: how the concepts gender and tense are written into the activation h; the network is the same, only the inputs change.
- Edit: add to h a concept’s direction found on training data; compose: applied together, two edits make both intended changes.
- Arrows and table: how far the edits move each concept, where 1 is the gender edit’s effect on training data; dashed ring: the change wanted.
- Toy values: two concepts, a linear mixing, and new inputs that turn it by the slider’s angle.