MNIST-Handwritten-Digit-Recognition-with-CNN-in-Python-using-Keras

This project implements and explains Python code to recognize handwritten digits (MNIST dataset) with a CNN using Keras.

Objective

• Understand how CNN helps by implementing first with MLP and then moving onto CNN
• Get an idea about CNN implementation in Python / Keras

Gratitude

I am reading the book 'Deep Learning with Python' by Francois Chollet. I sincerely express my gratitude to the author for the simplicity in explanation and explanation with code for MLP/CNN based solution for Handwritten Digit Recognition using MNIST dataset.

I understood the code from the book and thought to write it myself in this notebook to build my understanding and intuition. I thought this way I can capture my inferences as I am reading the book and get an idea about how CNN works.

Key Design Considerations

This is a Multi-Class Classification problem (10 classes)

• Language: Python
• Deep Learning Package: Keras
• Dataset: MNIST dataset available with Keras
• Model: MLP, CNN

Key Results

MLP based solution

• Input (784x1 column vector) -> 1 hidden layer (512 neurons, relu) -> Output (10 classes, softmax)
• Number of parameters = 407050
• Training Accuracy = ~98.9%
• Test Accuracy = ~98.0%

CNN based solution

• Input (28x28x1 matrix) -> Conv(32 3x3 filters, relu) -> Maxpool(2x2) -> Conv(64 3x3 filters, relu) -> Maxpool(2x2) -> Conv(64 3x3 filters, relu) -> Flatten to a column vector for FCN -> FCN (64 outputs, relu) -> FCN (10 outputs, softmax)
• Number of parameters = 93322
• Training Accuracy = ~99.4%
• Test Accuracy = ~99.3%

Conclusion

CNN has the advantage of using the spatial information in the image as compared to MLP, hence achieves better accuracy with reduced parameters.