Experiment with evolving small neural networks to play games instead of training them with gradients.
Study a tree structured evolutionary approach to predicting text one character or word at a time.
Explore evolution based image denoising as an alternative to standard diffusion model training.
Browse and compare past training runs through a built-in dashboard with metrics and history.
| a1cst/genreg-lab | 0xallam/my-recipe | 0xhassaan/nn-from-scratch | |
|---|---|---|---|
| Stars | 0 | — | 0 |
| Language | Python | Python | Python |
| Last pushed | — | 2022-11-22 | — |
| Maintenance | — | Dormant | — |
| Setup difficulty | hard | moderate | moderate |
| Complexity | 5/5 | 2/5 | 4/5 |
| Audience | researcher | general | developer |
Figures from each repo's GitHub metadata at analysis time.
Requires Windows for the built-in terminal feature, plus Python and its dependencies, a GPU speeds up training but is not required.
GENREG LAB is a research project for training AI models using evolution instead of the usual gradient based training methods most machine learning uses today. Rather than adjusting a model gradually using calculus based techniques, this project starts with a population of small models, tests them, and repeatedly selects and mutates the better performing ones over many generations, similar to how natural selection works in biology. No pretrained weights or gradient descent are used anywhere in the project. The lab is built as a single web application with several experiment pages you can open in a browser. One page evolves small neural networks to play games like Snake, 2048, and CartPole under various survival pressures such as limited energy or scarcity. Another page called Tree LM evolves a tree shaped structure of tiny models that predict text one character or word at a time. A third experiment applies the same evolutionary approach to image denoising, a technique related to how some AI image generators work. There is also a page for visually assembling models by dragging and connecting nodes, and a dashboard for browsing and comparing past training runs. A significant part of the project is a staged effort to build a text prediction model entirely through evolution, described as progressing through separate stages such as a basic recurrent memory component, an attention mechanism for selectively recalling earlier text, and a technique for training the model on its own generated output to improve real world performance. The README reports specific numeric results from these experiments as evidence of progress. The application also includes real interactive terminal windows built into the browser interface, backed by a separate background process so that terminals and any programs running in them keep working even if the main web server restarts. Running the project requires Windows, Python, and the listed dependencies, and it works with or without an NVIDIA GPU, though the original experiments used one. It only runs on your local computer and is not exposed to the network by default. The project is licensed under AGPL-3.0, a copyleft license that generally requires sharing source code changes, including when run as a network service.
A research lab for training AI models purely through evolution and selection, with no gradient descent or pretrained weights.
Mainly Python. The stack also includes Python, Flask, WebSocket.
You can use and modify this software, but if you distribute it or run it as a network service, you must share your source code under the same license.
Setup difficulty is rated hard, with roughly 1h+ to a first successful run.
Mainly researcher.
This repo across BitVibe Labs
double-check against the repo, no cap.