Digit Recognizer

Model to recognize digits with a convolutional neural network, consider LeNet-5 Architecture with some improvements viz (two stacked 3x3 filters replace the single 5x5 filters, max pooling layers inplaces of average pooling layers, ReLU activation replaces sigmoid, Batch normalization and Dropout is added)

Code and Resources Used

• Tensorflow
• Numpy
• Matplotlib
• Python Version:3.6

LeNet-5 Architecture

Some improvements in LeNet-5 Architecture are:

• Two stacked 3x3 filters replace the single 5x5 filters.
• Max pooling layers inplaces of Average pooling layers.
• ReLU activation replaces sigmoid.
• Batch normalization is added
• Dropout is added

Model Accuracy

• Train Accuracy: 0.99525
• Test Accuracy: 0.9894