pngwn/linearizer-minimal-repro
Minimal Repro: "Who Said Neural Networks Aren't Linear?" (the Linearizer, arXiv:2510.08570) A ~250-line 2D toy that proves the paper's core principle end to end. Not a full-paper reproduction (no LPIPS, no image diffusion, no large-scale training) — just the smallest experiment that demonstrates: Induced linearity is exact. f(x) = g⁻¹(A·g(x)) satisfies f(a·x+b) = a·f(x) + b·f(0) to machine precision when the identity is evaluated in the space where the algebra happens (g-space)… See the full description on the dataset page: https://huggingface.co/datasets/pngwn/linearizer-minimal-repro.
Remove failed chunked-upload debris; superseded by linearizer_toy.png
Remove failed chunked-upload debris; superseded by linearizer_toy.png
Remove failed chunked-upload debris; superseded by linearizer_toy.png
Fix file list to match repo contents
Remove failed chunked-upload debris; superseded by linearizer_toy.png
figure b64 part 5/5
Run metrics: collapse, linearity, mode coverage
README
figure b64 part 4/5
Minimal Linearizer repro: final script with corrected induced-linearity check
figure b64 part 3/5
figure b64 part 2/5
Remove stray partial-upload artifact
Final repro script (corrected check 1, dtype fix, zero-init, ActNorm direction fix)
figure b64 part 1/5 (76-char wrapped)
Figure: data vs 100-step Euler vs one-step collapsed samples
figure b64 part 1/3
Results figure: data vs 100-step Euler vs one-step collapsed samples
Full repro script: train + induced-linearity/collapse checks + figure
Final metrics from run 2 (8000 steps, lr 1e-3)
Add README with results summary
initial commit