Improve network component and use LR annealing technique.

[nik1806]

This issue is to encourage use of recent advance and monitor performance change.

1. Using activation function SELU instead of RELU. Reaserch about the advantage and apply.
2. For training, we can use LR annealing. Perhaps, with a manual approach at the start.

Apply them individually and then in combination. Make sure to note all the performance changes. They could of use in research paper.

[nik1806]

For activation function, also look into SWISH, proposed by Google Brain and claimed that it performs better than ReLU on various datasets.