README.md

metadata

license: mit
tags:
  - automl
  - genetic-programming
  - loss-functions
  - neural-networks
  - bittensor
datasets:
  - synapz/automl-genes

AutoML Evolved Loss Functions

This repository contains evolved neural network loss functions discovered through distributed genetic programming on the Bittensor network (subnet 49 Hivetrain AutoML).

Overview

The genes stored here represent novel loss functions optimized for:

• Image classification tasks
• Neural network training efficiency
• Improved convergence rates

Repository Structure

/automl-genes ├── losses/ # Evolved loss function implementations ├── metrics/ # Performance metrics and evaluations └── metadata/ # Gene genealogy and evolution data

Technical Details

Loss functions are evolved using:

• Genetic programming with population size 100
• Tournament selection (size 3)
• Multi-architecture validation across:
  • MLP
  • ResNet
  • MobileNet V3
  • EfficientNet V2

Usage

The evolved loss functions can be imported and used as drop-in replacements for standard PyTorch loss functions in deep learning projects.

Project

Part of the Hivetrain AutoML subnet focused on discovering improved neural network components through distributed evolution.

For more information, visit:

• DistributedAutoML Repository
• Hivetrain Discord

---