Figure viewer
← Back to article
One scalar weight moves the two activation changes along the descending line (δw,−δw). The equal-positive request (+d,+d) is visibly off that line.
A shared-weight constraint · Exact mathematical model

One weight moves the two inputs in opposite directions, so it cannot shift both positively. This exact counterexample assumes a bias-free linear model; an editable bias changes what is possible.

Settings, scope and source files
Exact counterexample hw(x)=wxh_w(x)=wx at x=±1x=\pm1. A shared weight update yields (δw,−δw)(\delta w,-\delta w), so the direct request (+d,+d)(+d,+d) is unreachable for d≠0d\ne0. An editable bias would change this conclusion.
Plays once when in view · pause at any time