Python script for Feedforward ANN on MNIST hand-written digits dataset

Training set --> 60000 images

Test set --> 10000 images

Hyperparameters:

1. No of dense layers : 1
2. Activation functions : ReLU --> Softmax
3. Dropout percent (for regularisation) --> 20%
4. Optimizer : SGD
5. Loss function = Sparse categorical crossentropy
6. Epochs: 500 (Validation error is seen to plateau around 130epochs --> probable overfitting)

FINAL LOSS:

1. Train set --> 0.0109
2. Validation set --> 0.0667
3. Test set --> 0.0666

FINAL ACCURACY:

1. Train set --> 99.69%
2. Validation set --> 98.16%
3. Test set --> 98.16%

===>> This gap between training accuracy and test accuracy represents overfitting

===>> Confusion Matrix says that most misclasified labels were 2-->7 and 3-->5