Figure viewer
← Back to article
A weight-space step maps through J to a blue activation-space step. A shrinking Fisher ellipse centered on the desired change first touches the reachable line at that same endpoint.
Metric projection · Exact mathematical model

With a complete interface and its pulled-back Fisher metric, the two routes agree locally: move the weights along their natural gradient, or project the activation natural gradient onto reachable changes.

Settings, scope and source files
Exact tangent model: F=(2.5.51)F=\begin{pmatrix}2&.5\\.5&1\end{pmatrix}, v=(.8,1.4)v=(.8,1.4), and permitted changes u2=0u_2=0. Contours show 12(u−v)⊤F(u−v)\tfrac12(u-v)^\top F(u-v); first contact with the blue line gives the Fisher projection (1.15,0)(1.15,0), distinct from the Euclidean projection (.8,0)(.8,0). With J=(1,0)⊤J=(1,0)^\top, the weight metric is Gθ=J⊤FJ=2G_\theta=J^\top FJ=2. For the linear objective with differential Fv, the weight natural gradient pushed through J equals that projection. The two arrows grow together to display this identity. This is an analytic construction, not Qwen measurements or an optimization record.
Plays once when in view · pause at any time