Collatz Conjecture AI Research System

Advanced AI system for analyzing the Collatz Conjecture using Deep Learning and parallel brute-force search

🎯 Overview

Key Features

• 🧠 Transformer-based Neural Network for sequence prediction
• 🔍 Multi-threaded C++ Loop Searcher (Floyd's Cycle Detection)
• ⚡ Native C++ Data Engine for maximum performance
• 📊 Real-time Discord Integration for monitoring
• 🎓 Curriculum Learning with "Hard Mode" candidates (n > 2^68)
• 🔬 Advanced Optimizations: Mixed Precision (AMP), Cosine Annealing LR, Gradient Clipping

🚀 Quick Start

Prerequisites

• GPU: NVIDIA GPU with 6GB+ VRAM (tested on RTX 3070 Ti)
• CPU: Multi-core processor (tested on Ryzen 5900X)
• RAM: 16GB+ recommended
• OS: Linux (tested on Arch Linux)
• Python: 3.13+
• CUDA: 12.8+

Installation

Usage

🏗️ Architecture

AI Model (GPU)

Input: Parity Vector [0, 1, 0, 1, 1, ...]
  ↓
Embedding Layer (3 → 128d)
  ↓
Positional Encoding
  ↓
Transformer Encoder (4 layers, 4 heads)
  ↓
Dual Heads:
  ├─ Stopping Time Prediction (Regression, Log-Space)
  └─ Next Step Prediction (Classification)

• Model Size: 128d, 4 layers, 4 attention heads
• Batch Size: 512
• Optimizer: AdamW with Cosine Annealing
• Loss: Huber Loss (stopping time) + CrossEntropy (sequence)

Loop Searcher (CPU)

• Multi-threaded C++ implementation
• 128-bit integer support (__int128)
• Detects non-trivial cycles
• Runs in background during training

📊 Performance

Training Metrics (100k steps)

Loop Search Results

• Numbers Checked: 22,000,000 per run
• Range: [2^68, 2^68 + 22M]
• Non-trivial Cycles Found: 0 (as expected)

🛠️ Technical Details

Optimizations

1. Mixed Precision Training (AMP)
   • Reduces VRAM usage by ~40%
   • Increases training speed by ~30%
2. Native C++ Engine
   • 20-30% faster data generation
   • Supports numbers > 2^64 (128-bit)
3. Curriculum Learning
   • 50% "Normal" data (sequential numbers)
   • 50% "Hard" data (n > 2^68, special patterns)
4. Learning Rate Scheduling
   • Cosine Annealing: 1e-4 → 1e-6
   • Smooth convergence, prevents oscillation

File Structure

collatz_ai/
├── src/
│   ├── train.py              # Main training script
│   ├── model.py              # Transformer architecture
│   ├── engine.py             # Numba-optimized data generation
│   ├── dataset.py            # PyTorch Dataset/DataLoader
│   ├── analyze.py            # Model analysis & visualization
│   ├── discord_bot.py        # Discord webhook integration
│   ├── collatz_core.cpp      # C++ data engine
│   ├── loop_searcher.cpp     # C++ parallel loop searcher
│   ├── native_engine.py      # Python bindings (ctypes)
│   └── loop_search.py        # Loop searcher wrapper
├── checkpoints/              # Model checkpoints (auto-saved)
├── requirements.txt          # Python dependencies
├── run.sh                    # Setup & run script
└── README.md                 # This file

📈 Results & Insights

What the Model Learned

1. Stopping Time Prediction: Near-perfect accuracy (99.97%)
2. Parity Patterns: Strong recognition of even/odd sequences
3. Anomaly Detection: Identifies numbers with unusual stopping times

Known Anomalies

Number: 1249, Actual: 176, Predicted: 233, Error: 57
Number: 1695, Actual: 179, Predicted: 236, Error: 57
Number: 1742, Actual: 179, Predicted: 235, Error: 56

🔬 Research Applications

For Mathematicians

• Analyze stopping time distributions
• Identify exceptional numbers
• Visualize sequence embeddings (PCA)

For ML Researchers

• Benchmark for sequence prediction
• Study curriculum learning effects
• Explore transformer behavior on mathematical sequences

For Collatz Enthusiasts

• Automated large-scale verification
• Pattern discovery in high ranges (> 2^68)
• Real-time progress monitoring via Discord

🚧 Future Work

• Distributed training across multiple GPUs
• Larger model (256d, 6 layers) from scratch
• GPU-accelerated loop detection
• Extended search range (2^100 - 2^120)
• Hybrid LSTM+Transformer architecture

📝 Configuration

🤝 Contributing

• Model architecture improvements
• Faster loop detection algorithms
• Better anomaly detection
• Visualization enhancements

📄 License

• Use for any purpose (personal, commercial, research)
• Modify and improve the code
• Distribute original or modified versions
• Sell modified versions

• Share source code of any modifications
• Use the same GPL v3 license
• State significant changes
• Include copyright and license notices

🙏 Acknowledgments

• Collatz Conjecture: Lothar Collatz (1937)
• PyTorch Team: For the amazing framework
• Numba Team: For JIT compilation magic
• Community: For mathematical insights

📧 Contact

• Issues: GitHub Issues
• Discussions: GitHub Discussions