🧪 ML-Research: Experiments in Machine Learning & AI

Welcome to ML-Research, a public repository of research-grade machine learning experiments, prototypes, and conceptual studies.
This repo is designed as a sandbox for testing new ideas, validating theories, and producing reproducible results across various ML domains.

---

🔬 Project Goals

• 📚 Document machine learning experiments with clarity and rigor
• 🧠 Explore new model designs and architectural modifications
• ⚙️ Evaluate practical and theoretical performance tradeoffs
• 📈 Provide reproducible and well-structured notebooks for community use

---

📁 Current Experiments

mnist_notebook_00.ipynb

Minimalist CNN for General-Purpose Experimentation

This notebook implements a clean, performant convolutional neural network (CNN) for the MNIST digit classification problem, achieving ~99% accuracy with minimal complexity.

🔍 Features:

• ✅ Solid MNIST benchmark (~98.9–99.0% test acc)
• 🧩 Modular design for easy layer or logic injection
• ➕ Optional integration of synthetic data via included mixing cell
• 📉 Loss curves and batch-wise metrics for quick validation cycles
• 🧪 Ideal for rapid iteration and reproducible research

---

mnist_synthetic_00.ipynb

Houdini-Generated Synthetic Dataset Exploration

Explores a procedurally generated digit dataset created in Houdini and exported for ML purposes.

🧪 Purpose: Measure impact of synthesized training data on generalization accuracy.

📊 Findings: Adding ~40,960 images (~41% of training data) improved test accuracy to ~99.1%, compared to ~98.9% with MNIST-only training.

---

🏗 Houdini Data Synthesis Pipeline

Scene: ./houdini/digit_gen_v02.hip

Overview:

• Renders 64 frames of 8×8 digit sprite sheets to $HIP/render
• Cuts them into 28×28 grayscale images using compositing + TOPs
• Outputs to $HIP/output/{0..9} folder structure for label consistency

This procedural pipeline enables data generation in low-data domains and can be adapted to RGB, animation, 3D points, textures, etc.

---

💾 Models

• mnist_cnn.pth — Trained on MNIST only (5 epochs)
• mnist_cnn_01.pth — Trained on MNIST only (15 epochs)
• mnist_cnn_02.pth — Trained on MNIST + synthetic digits (enhanced result)

---

📦 Dependencies

All experiments use Python ≥3.12 and are tested with:

pip install torch torchvision matplotlib

(Get Torch Here)

🔖 License This repository is open-sourced under the MIT License.