The Dirty Man — self-reconfiguring computation
A self-reconfiguring neural architecture that rewires which network computes each sample, driven by what it sees and what it wants. Core component: the Switch Operator, a router over nine neural primitives (linear, dense, ReLU, CNN, RNN, LSTM, GAN, autoencoder, transformer), each with its own inductive lens.
Headline results (full scale, T4 GPU)
| Result | Numbers |
|---|---|
| Mixed-domain (Protocol A) | one operator 0.834 beats every fixed network it contains |
| Real handwriting, zero synthetic overlap (Protocol E) | switch 0.334 > static CNN 0.313 > static MLP 0.312; router identifies real digits need spatial lenses (CNN 0.67) |
| Training-time intervention | energy violation cut 16x (0.42 → 0.026) by routing to a physics expert |
| Flagship (no single net can obey two laws) | routed energy error up to 336x lower than a static net |
| Discovered-law flagship (no physics hardcoded) | learned specialists + router that discovers the law: 5–720x better than a single static net; scaling 4.5–9x better at every law count |
Theory
Five theorems with proofs in the paper, including a single-map separation bound (any fixed map must mis-represent at least one law by half the per-step energy gap) and a scaling bound (a single map's error is bounded below by a constant independent of the number of laws, while routing's error decays with router accuracy).
Reproduce
pip install -r requirements.txt
python run_experiments.py --smoke # ~1 min sanity check
python flagship_regime_routing.py # flagship (hardcoded harnesses)
python flagship_discovered_law.py # discovered-law flagship + scaling
Full source, paper (NMI-style + IEEE), figures, and committed results: https://github.com/sehajr-singhs/dirty-man
Website: https://sehajr-singhs.github.io/dirty-man/
Kaggle GPU kernel: https://www.kaggle.com/code/sehajrsingh/dirty-man-headline-experiments