This python script uses Keras based on Theano backend to train for categorization of MNIST data by application of a neural network consisting of 3 layers-
1. Input layer accepting digits of MNIST dataset, having shape (784,).
2. Hidden layer with 350 neurons.
3. Output layer with 10 neurons - representing the 10 output classes (digits) for MNIST dataset.
The network has been applied on half of MNIST dataset (- a collection of 42000 handwritten digit (0-9) images) for quick computation with a quarter of this dataset used for validation. Since the dataset is too heavy to be uploaded on github (76.8 MB), it can be found at https://goo.gl/Wyl4hX.
The output of the python script can be found in the results.txt file.
After 50 epochs, the training accuracy was 94.13% while validation accuracy was 92.13%. The model seemed to have slightly overfit as an accuracy of 93.16% was achieved after 49 epochs.
Accuracy Progression:
(Blue: Training Accuracy Red: Validation Accuracy)
Loss Progression:
(Blue: Training Loss Red: Validation Loss)
