Lightweight MNIST CNN (4.5K Parameters)
This is an ultra-lightweight Convolutional Neural Network (CNN) with only 4.5K parameters, trained on the classic MNIST dataset for handwritten digit recognition.
Despite its minimal size, the model achieves 94.48% accuracy on the test split, making it highly efficient for edge devices or CPU-only environments.
The full training workflow and source code can be found on GitHub and weights can be found on Hugging Face
Model Performance
- Task: Image Classification (Handwritten Digits 0-9)
- Dataset: MNIST
- Test Accuracy: 94.48%
- Parameter Count: ~4,500
Quick Start / Usage
You can load and use this model directly using the code in github repository