Hfrxdjt

Keras gives you the option to apply regularization differently to different layers. I mean, why not? Though when I first learned about neural nets (from ESL), I thought of it as a global parameter.

Global is simpler to tune, but obviously a global penalty can be no better than equally efficient when compared to some optimal set of layerwise ones.

So, what are the cases where different penalties for different layers will work better than a single global penalty, and better-enough to be worth the bother?

Regularisation is a technique to solve overfitting.

This feature from Keras is going to help a lot in many scenarios. Few times we don't want simpler but a granulr tuning.

1. CNN: we all know that each convolution layer can contribute to certain set of features from the dataset, and we now a days know what it is trying to do, by defining regularisation to each layer differently, we can better understand how each layer is effecting the final output


2. Transfer Learning: Where we want learn from the already trained network, and use that domain knowledge. now during this, we can now control, how much we want to regularise before/after merging from the base network.


3. Multi Task Learning: This is a technique in which we learn multiple tasks together, now with this kind of regularisation we can now control before the merge of the layers, how much of the information can be merged.

these are the quick things i could think of. But there are definitely lots of other uses.

Vote up, if this helps ;)