Quantum Neural Horror
a neuroevolution sandbox that rewrites its own topology
Written to find out whether the internals were understood, or only the diagrams.
Backpropagation is not the only way to search for an architecture, and this is the other way, run rather than described. The network modifies its own topology while training, adding and removing neurons, and its activation functions are selected genetically rather than chosen up front. Simulated annealing with Metropolis-Hastings acceptance lets it take a worse step to escape a local minimum, which is the whole reason a topology search is not simply hill climbing.
The design constraint is that it all fits in one file and draws itself. Five live matplotlib panels show the search as it happens rather than after it: the architecture as it changes shape, the loss, the mutation history, and the complexity measures the run tracks about itself. Watching an architecture search is a different kind of understanding from reading its final accuracy, and that is the point of the project rather than a decoration on it.
Underneath the presentation it is ordinary, careful machine-learning engineering. L2 and complexity penalties keep the topology search from simply growing the network until it memorises the data, which is the failure mode a self-modifying architecture walks into first. It is a teaching object and a sandbox, not a model anyone should deploy, and the repository is clear about that.
- Python 100%
Built from
What it is for
- A sandbox and a demonstration, not a production trainer. It is small on purpose and reads end to end.
- The consciousness and meta-cognition panels are named as a joke in the repository. They are complexity measures over the evolving topology, not a claim about the network.
Screenshots
Roads out
Related articles
- Neuroevolution: networks that grow instead of learn
Almost every neural network is trained by backpropagation. There is another way: let the network mutate its own structure and keep whatever works. Quantum Neural Horror does that, and lets you watch it happen.
4 min read · 2026-08-28
Neighbouring holdings
- Read the charter of NCAA Outcome Predictor shares machine-learning, python
- Read the charter of The Pulse Engine shares python