Figure 2. A 3D cartoon of linear steering. Each dot is an activation vector h for one input. The steering vector v is the difference between the two cluster means (difference-in-means, as in CAA). The grey plane is where later layers switch from reading the input as negative to positive. Adding α·v to a negative example moves it across this plane, and later layers then produce a positive output. Real activation vectors have thousands of dimensions; this picture keeps three.