Python script for Feedforward ANN on MNIST hand-written digits dataset
Training set --> 60000 images
Test set --> 10000 images
Hyperparameters:
- No of dense layers : 1
- Activation functions : ReLU --> Softmax
- Dropout percent (for regularisation) --> 20%
- Optimizer : SGD
- Loss function = Sparse categorical crossentropy
- Epochs: 500 (Validation error is seen to plateau around 130epochs --> probable overfitting)
FINAL LOSS:
- Train set --> 0.0109
- Validation set --> 0.0667
- Test set --> 0.0666
FINAL ACCURACY:
- Train set --> 99.69%
- Validation set --> 98.16%
- 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