This project implements the LeNet-5 architecture from scratch in Python, using only numpy.
This version of LeNet-5 has the same basic structure as the original, but some details were changed to reflect more modern sensibilities (e.g., ReLU activation).
This modern version manages to crack 99.3% accuracy against the MNIST dataset; the original version reported an accuracy of 99.05%.
In the Jupyter notebook, we also try some classical ML models and a few simpler deep learning architectures before using LeNet-5.
This project also contains code to load the MNIST dataset, do some basic preprocessing, and score predictions. It also provides some useful visualizations to show the overall performance of a model and sample both correct and incorrect predictions.
In the real world, a deep learning framework such as Keras would be used instead. However, some experts recommend implementing deep learning yourself to understand how it works—and how frameworks like Keras work.