My Kaggle.com entry for MNIST classifier. This one got me 0.99657 accuracy.
- 2 CNNs, with 16 5x5 pixel filters, ReLU activations.
- MaxPooling - 2x2 pool size with 2x2 strides
- Fully connected layer, 128 neurons, ReLU activations, 0.5 dropout for regularization
- final fully connected output layer, softmax activations.
- loss function is categorical cross entropy
- GD optimizer Adam