Property-Driven Machine Learning

A general framework for property-driven machine learning that enables incorporating logical properties and constraints into neural network training and evaluation.

Features

• Multiple Logic Systems: Support for Boolean logic, fuzzy logics (Gödel, Łukasiewicz, Reichenbach, Yager, etc.), Signal Temporal Logic (STL), and DL2
• Property Constraints: Built-in constraint classes for robustness properties, Lipschitz constraints, output bounds, and group fairness
• Adversarial Training: PGD and Auto-PGD attack implementations for constraint evaluation
• Gradient Normalization: GradNorm for balancing multiple training objectives

Installation

From PyPI (Recommended)

# Install the latest version from PyPI
pip install property-driven-ml

# Or install with uv (faster)
uv add property-driven-ml

From Source (Development)

For development or to run the latest features:

Prerequisites

This project uses uv for fast Python package and project management. Install uv first:

# On macOS and Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

# On Windows
powershell -c "irm https://astral.sh/uv/install.ps1 | iex"

# Or with pip
pip install uv

Installation Steps

# Clone the repository
git clone https://github.com/tflinkow/property-driven-ml.git
cd property-driven-ml

# Install dependencies and the package (recommended)
uv sync

# Or install in development mode with pip
pip install -e .

# Or install directly
pip install .

Requirements

• Python 3.11+
• PyTorch 2.5.1+
• CUDA support (optional, for GPU acceleration)

Note:

• For PyPI installation: Use standard python commands
• For local development with uv: Use uv run python to run scripts with the managed environment

Quick Start

Using the Command Line Interface

After installation from PyPI:

# Get help on available options
property-driven-ml --help

# Run a training experiment
property-driven-ml \
  --data-set=mnist \
  --batch-size=64 \
  --lr=0.001 \
  --epochs=10 \
  --input-region="EpsilonBall(eps=0.1)" \
  --output-constraint="StandardRobustness(delta=0.05)" \
  --experiment-name="mnist_robustness" \
  --logic=GD

For local development:

# Run from the repository root
uv run python main.py --help

Basic API Usage

import torch
import property_driven_ml.logics as logics
import property_driven_ml.constraints as constraints
import property_driven_ml.training as training

# Create a logic system
logic = logics.GoedelFuzzyLogic()

# Define a robustness constraint
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
constraint = constraints.StandardRobustnessConstraint(device, delta=0.1)

# Set up adversarial attack for constraint evaluation
x_sample = torch.randn(1, 3, 32, 32).to(device)  # Example input
oracle = training.PGD(x_sample, logic, device, steps=20, restarts=10, step_size=0.01)

# Use in training loop
model = torch.nn.Sequential(
    torch.nn.Flatten(),
    torch.nn.Linear(3072, 128),
    torch.nn.ReLU(),
    torch.nn.Linear(128, 10)
).to(device)

# Evaluate constraint satisfaction
x = torch.randn(8, 3, 32, 32).to(device)
x_adv = oracle.attack(model, x, None, x.min(), x.max(), constraint)
loss, satisfaction = constraint.eval(model, x, x_adv, None, logic)

Available Logic Systems

import property_driven_ml.logics as logics

# Access logic systems through the logics module
logic_boolean = logics.BooleanLogic()                    # Classical Boolean logic
logic_godel = logics.GoedelFuzzyLogic()                  # Gödel fuzzy logic
logic_lukasiewicz = logics.LukasiewiczFuzzyLogic()       # Łukasiewicz fuzzy logic
logic_reichenbach = logics.ReichenbachFuzzyLogic()       # Reichenbach fuzzy logic
logic_yager = logics.YagerFuzzyLogic()                   # Yager fuzzy logic
logic_stl = logics.STL()                                 # Signal Temporal Logic
logic_dl2 = logics.DL2()                                 # DL2 logic

Available Constraints

import property_driven_ml.constraints as constraints

# Access constraint classes through the constraints module
robustness = constraints.StandardRobustnessConstraint(device, delta=0.1)    # Adversarial robustness
lipschitz = constraints.LipschitzRobustnessConstraint(device, L=1.0)        # Lipschitz continuity
output_bounds = constraints.AlsomitraOutputConstraint(device, lo, hi)       # Output bounds
group_fair = constraints.GroupConstraint(device, groups, delta=0.1)         # Group fairness
input_region = constraints.EpsilonBall(dataset, eps=0.1, mean, std)         # Input perturbation regions

Architecture

The framework is organized into several key modules:

• property_driven_ml.logics: Logic systems for constraint evaluation
• property_driven_ml.constraints: Property constraint definitions
• property_driven_ml.training: Training utilities (attacks, gradient normalization)
• property_driven_ml.util: Utility functions

Examples

Command-Line Usage

If you installed from PyPI, use the CLI directly:

# Run a complete training experiment
property-driven-ml \
  --data-set=mnist \
  --batch-size=64 \
  --lr=0.001 \
  --epochs=10 \
  --input-region="EpsilonBall(eps=0.1)" \
  --output-constraint="StandardRobustness(delta=0.05)" \
  --experiment-name="mnist_robustness" \
  --logic=GD

For local development:

# Run from the root directory with uv
uv run python main.py \
  --data-set=mnist \
  --batch-size=64 \
  --lr=0.001 \
  --epochs=10 \
  --input-region="EpsilonBall(eps=0.1)" \
  --output-constraint="StandardRobustness(delta=0.05)" \
  --experiment-name="mnist_robustness" \
  --logic=GD

# Or run the shell script with multiple experiments
cd examples && bash run.sh

Programmatic Usage

Write your own scripts using the framework:

import property_driven_ml.logics as logics
import property_driven_ml.constraints as constraints

# Your custom training code here...

For local development examples:

# Simple API demonstration
uv run python examples/simple_example.py

Available Examples (Local Development)

• main.py: Full training pipeline with command-line interface (root level)
• examples/simple_example.py: Clean API demonstration
• examples/models.py: Neural network architectures for different tasks
• examples/alsomitra_dataset.py: Custom dataset implementation
• examples/run.sh: Batch training experiments

Contributing

Contributions are welcome! Please feel free to submit pull requests or open issues for bugs and feature requests.

Development Setup

This project uses uv for dependency management and development workflows:

# Clone and set up the development environment
git clone https://github.com/tflinkow/property-driven-ml.git
cd property-driven-ml

# Install all dependencies (including dev dependencies)
uv sync

# Run tests
uv run pytest

# Run examples and scripts
uv run python main.py --help
uv run python examples/simple_example.py

# Add new dependencies
uv add torch torchvision  # Runtime dependency
uv add --dev pytest mypy  # Development dependency

License

This project is licensed under the MIT License - see the LICENSE file for details.

Citation

If you use this framework in your research, please cite:

@article{flinkow2025general,
  title={A General Framework for Property-Driven Machine Learning},
  author={Flinkow, Thomas and Casadio, Marco and Kessler, Colin and Monahan, Rosemary and Komendantskaya, Ekaterina},
  journal={arXiv preprint arXiv:2505.00466},
  year={2025},
  url={https://arxiv.org/abs/2505.00466}
}

You can also cite the software implementation:

@software{property_driven_ml,
  title={Property-Driven Machine Learning Framework},
  author={Thomas Flinkow},
  year={2025},
  url={https://github.com/tflinkow/property-driven-ml}
}