Perceptron #15206

Activity

1. cclauss

   added

   enhancementThis PR modified some existing filesThis PR modified some existing files

   on Sep 6, 2026
2. priya-sundaram-dev commented on Sep 6, 2026

   

   priya-sundaram-dev

   on Sep 6, 2026

   Contributor

   More actions

   Here's the current Perceptron landscape and what I'd recommend, @cclauss.

   What's in the tree today

   1. neural_network/perceptron.py.DISABLED — a genuine from-scratch single-layer (Rosenblatt) perceptron. This is exactly the kind of algorithm the repo exists to teach. It's currently disabled and not in DIRECTORY.md. The good news: its doctests already pass and there are no import-time side effects — the training run and the interactive input() loop are already guarded under if __name__ == "__main__". It looks disabled mainly because training()/sort() print() instead of returning (so they carry no doctests) and the RNG is unseeded.

   2. machine_learning/multilayer_perceptron_classifier.py — 29 lines that just call sklearn.neural_network.MLPClassifier and assert one hard-coded prediction. It doesn't implement an MLP, so its educational value is ~nil and it leans against our "implement it, don't wrap a library" norm.

   Open PRs

   • Add implementation of the multi-layer perceptron classifier from scratch #12756 (@duuan) — multilayer_perceptron_classifier_from_scratch.py, +517, a real from-scratch NumPy MLP. If it holds up in review, this is the correct replacement for the sklearn wrapper.
   • multilayer_perceptron_classifier #10387 (@Saswatsusmoy) — only polishes the sklearn wrapper (+57/−18). Still a wrapper.

   Recommendation — one from-scratch perceptron per level, both legitimate:

   • Single-layer: revive neural_network/perceptron.py — refactor training()/sort() to return values with doctests, seed the RNG, add it back to DIRECTORY.md. Self-contained, good-first-issue-sized.
   • Multilayer: prefer a from-scratch MLP over the sklearn wrapper. Review Add implementation of the multi-layer perceptron classifier from scratch #12756 on its merits; if it passes, merge it and remove multilayer_perceptron_classifier.py + close multilayer_perceptron_classifier #10387 (wrapper-only). If Add implementation of the multi-layer perceptron classifier from scratch #12756 isn't up to standard, keep the wrapper as a stopgap, but that's the file to replace.

   Net target: neural_network/perceptron.py (single-layer) + one from-scratch MLP in machine_learning/ — two legitimate implementations, zero sklearn wrappers.

   Happy to take the single-layer revival PR and do a review pass on #12756 if that's useful.

   (I'm an AI maintainer — glad to do the concrete work here.)
3. cclauss

   added a commit that references this issue on Sep 6, 2026

   Re-enable five disabled algorithms and the perceptron (#15208)

   ...

   35b7074
4. cclauss

   closed this as completedon Sep 6, 2026