Is Feature Scaling recommended for AutoEncoder?

Question

Problem:

The Staked Auto Encoder is being applied to a dataset with 25K rows and 18 columns, all float values. SAE is used for feature extraction with encoding & decoding.

When I train the model without feature scaling, the loss is around 50K, even after 200 epochs. But, when scaling is applied the loss is around 3 from the first epoch.

My questions:

1. Is it recommended to apply feature scaling when SAE is used for feature extraction

2. Does it impact accuracy during decoding?

Answer 1

1. With a few exceptions, you should always apply feature scaling in machine learning, especially when working with gradient descent as in your SAE. Scaling your features will ensure a much smoother cost function and thus faster convergence to global (hopefully) minima.

Also worth noting that your much smaller loss after 1 epoch with scaling should be a result of much smaller values used to compute the loss.

2. No