Discovering machine learning algorithms from scratch through evolutionary computation.
The future of AI starts here.
AutoML Zero is a revolutionary approach to machine learning that discovers algorithms from scratch, rather than just tuning existing ones. Think of it as "AI that creates AI."
Researchers manually choose algorithms (SVM, Random Forest, Neural Networks)
Optimize existing parameters within predefined algorithm structures
Small gains through better tuning and feature engineering
Evolutionary algorithms discover entirely new ML algorithms from basic operations
Starting with only basic mathematical operations (+, -, *, /, exp, log)
Potentially discovering algorithms humans haven't thought of yet
Custom DSL for representing algorithms with setup, predict, and learn phases. Enables evolutionary operations on algorithm structure.
Decentralized computation network for distributed algorithm evolution. Miners contribute compute power to the discovery process.
High-performance algorithm execution using NumPy arrays. Enables efficient evaluation of thousands of algorithm variants.
Starting with digit classification (0 vs 1) on downscaled MNIST. Perfect testbed for algorithm discovery with clear success metrics.
Multi-generational evolution with mutation, crossover, and selection. Algorithms improve over time through natural selection.
Live monitoring of algorithm evolution progress. Watch as new algorithms emerge and improve over generations.
We're currently focused on building and validating our AutoML Zero implementation. This foundational work will enable the full HiveTensor platform.
While we're building the foundation, there are several ways to get involved with our AutoML Zero research.
Check out our AutoML Zero implementation on GitHub and contribute to the development.
View RepositoryDive into AutoML Zero research papers and understand the theoretical foundations.
Read Documentation