Skip to content

fix: NonlinearFeaturesLayer on matrix input, EuSNCell allocations - #567

Merged
MartinuzziFrancesco merged 1 commit into
masterfrom
fm/fix-nfl-matrix-eusn-allocs
Oct 24, 2026
Merged

MartinuzziFrancesco merged 1 commit into
masterfrom
fm/fix-nfl-matrix-eusn-allocs

Conversation

@MartinuzziFrancesco

@MartinuzziFrancesco MartinuzziFrancesco commented Oct 24, 2026 •

Copy link
Copy Markdown
Collaborator

NonlinearFeaturesLayer only had an AbstractVector method. Reservoir cells inside a ReservoirChain output n x 1 matrices, so a chain with a NonlinearFeaturesLayer after a cell threw a MethodError in collectstates, train! and predict. Matrix input is now handled column by column, so non-elementwise features such as kron(x, x) work, and the Any[] buffer is gone. The vector method returns the same values as before.

EuSNCell built W - W' - gammaI plus a dense identity on every step. It now computes Wh - W'h - gammah (324 KB -> 5.6 KB per step at n = 200). The bias is added with the input term, after the recurrent difference, so a large symmetric W cannot round it away.

Checklist

  • Appropriate tests were added
  • Any code changes were done in a way that does not break public API
  • All documentation related to code changes were updated
  • The new code follows the
    contributor guidelines, in particular the SciML Style Guide and
    COLPRAC.
  • Any new documentation only uses public API

Additional context

Add any other context about the problem here.

…tions

NonlinearFeaturesLayer only had an AbstractVector method. Reservoir cells
inside a ReservoirChain output n x 1 matrices, so a chain with a
NonlinearFeaturesLayer after a cell threw a MethodError in collectstates,
train! and predict. Matrix input is now handled column by column, so
non-elementwise features such as kron(x, x) work, and the Any[] buffer is
gone. The vector method returns the same values as before.

EuSNCell built W - W' - gamma*I plus a dense identity on every step. It now
computes W*h - W'*h - gamma*h (324 KB -> 5.6 KB per step at n = 200). The
bias is added with the input term, after the recurrent difference, so a
large symmetric W cannot round it away.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@MartinuzziFrancesco
MartinuzziFrancesco merged commit 20f50fb into master Oct 24, 2026
60 checks passed
@MartinuzziFrancesco
MartinuzziFrancesco deleted the fm/fix-nfl-matrix-eusn-allocs branch October 24, 2026 14:06
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant