Manuscript Psy01-A@1 · Psy01-A.tex · example · 94477fa3

Example: An intervening update changes the final support

Stability, hard-mask discontinuity, and schedule dependence › Nested pruning and a quantitative learning-induced defect · lines 1462–1479

Let $w = (3, 2, 1)$ . Early retention of two coordinates gives $x = (3, 2, 0)$ . One gradient step of size one for

L (x) = \frac{1}{2} (x_{1} - \frac{1}{2})^{2}

gives $y = (1/2, 2, 0)$ . Late retention of one coordinate gives

H_{1} (y) = (0, 2, 0),

whereas one-shot pruning of the original vector gives

H_{1} (w) = (3, 0, 0) .

Their Euclidean distance is $13$ . This is learning-induced re-ranking, not noncommutation of fixed masks.

Displayed formulas in this passage