Why does we need to scale the input to [0,1]?

[hjung]

I've noticed that people often use transform_param to scale the input to [0, 1].

E.g. from http://caffe.berkeleyvision.org/tutorial/data.html,

  # common data transformations
  transform_param {
    # feature scaling coefficient: this maps the [0, 255] MNIST data to [0, 1]
    scale: 0.00390625
  }

Is there any advantage to scale the input range to [0, 1]?

It seems that this scaling strategy is not limited to deep learning, but can be applied to all the fields of machine learning. But, I can't find any reasonable reason.

Thanks for your reply in advance.