CNN Framework

This repository contains my own implementation of a Convolutional Neural Network (CNN) framework for model training and model creation.

Highlights

• Optimized Computations: Used dot products where applicable to enhance computational speed.
• Tested with LeNet-5: Successfully implemented the LeNet-5 architecture to validate the framework. Tested on MNIST dataset with preprocessing (96.5% Accuracy). Minor adjustments may be required to implement other CNN architectures.

Results

Download

Link to download the trained test model

Notes

• Partial Connections in Layer C3: I did not use partial connections for layer C3, as described in the original LeNet-5 paper, to simplify and speed up computations by avoiding nested loops.
• Classification Layer: The final layer uses Softmax for classification instead of the Radial Basis Function (RBF) described in the original paper.

Reference

LeCun et al.'s original paper on LeNet-5:
Gradient-Based Learning Applied to Document Recognition (1998)