Skip to content

Latest commit

 

History

History

nn_classification_of_handwritten_letters

README.md

🧠 EMNISTN: High-Performance Handwritten Letter Classification

PyTorch Status Accuracy


📌 Project Overview

Developed for the Codecademy ML Bootcamp Contest (Jan 2026), this project focused on high-efficiency Deep Learning. The challenge required building a neural network to classify handwritten letters (EMNIST) from scratch, strictly limited to <1M parameters.

My architecture, EMNISTNet, is a highly optimized CNN that achieves 94.87% test accuracy with only ~697k parameters. Key technical implementations included Label Smoothing for better generalization, Dropout for regularization, and advanced feature extraction layers. Adhering to production standards, the model prioritizes a high accuracy-to-parameter ratio, making it ideal for edge computing or resource-constrained environments.


🚀 Key Results

Metric Performance Notes
Test Accuracy 94.87% High generalization (Validation Acc: 95.26%)
Model Size 0.7 MB 697,594 Parameters (70% of budget)
Inference GPU/CPU Auto-detects hardware via predict.py

🛠️ Technical Approach & MLOps

To achieve top-tier performance within constraints, I implemented:

  • Architecture: Custom VGG-style CNN with Modular ConvBlocks (GELU Activations + BatchNorm).
  • Optimization: AdamW optimizer + OneCycleLR scheduler for superconvergence (15 epochs).
  • Data Engineering: Custom pipeline with FixOrientation, RandomRotation, and GaussianNoise for robustness.
  • Regularization: Heavy Dropout + Label Smoothing (0.1) to handle ambiguous letters (e.g., 'I' vs 'l').
  • Reproducibility: SHA256 integrity checks and automated model checkpointing.

📊 Visualizations

Training Dynamics

Training Curves

Confusion Matrix

Confusion Matrix


💻 How to Run

  1. Clone the repository. 
    git clone https://github.com/MiltonValleL/CodeCademy-DataScience-ML-Portfolio.git
  2. Install dependencies: 
    pip install -r requirements.txt
  3. Run inference: To test the model on random samples and verify the architecture: 
    python predict.py

Note: The script automatically detects if a GPU (CUDA) is available for acceleration.



🤝 Contact

I am a Data Science enthusiast passionate about Machine Learning and statistical modeling.

If you have any questions about this project or would like to connect, feel free to reach out!


Author: Milton Rodolfo Valle Lora

LinkedIn: Please click here