Difference between 'dropout' and 'backprop' arguements in script about dropout

[mdenil]

The code always builds a theano graph for both dropout and non-dropout versions of the network. The choice between dropout and no dropout at training time is made here https://github.com/mdenil/dropout/blob/master/mlp.py#L260 and here https://github.com/mdenil/dropout/blob/master/mlp.py#L314

For an explanation of dropout you should read the arxiv paper linked in the readme.

[saatvikshah]

I think I shouldve phrased the question better. Even in your backprop network you have shared weights with the Dropout Layers as given here https://github.com/mdenil/dropout/blob/master/mlp.py#L130. Is there some motive behind this? Normally in backprop hidden layers would have automatically initialized their own weights without passing the values of W and b manually from the dropout layer, Right?

[mdenil]

The dropout and non dropout layers share weights so that the same network can be evaluated both with and without dropout. If you compute dropout_cost then you get a forward pass with dropout applied, but if you compute cost then you get a forward pass through the same network with no dropout (and with appropriately scaled weights).

The computational graph looks like this:

  cost/errors    dropout_cost/dropout_errors
      |              |
HiddenLayers   DropoutHiddenLayers <--- these share weights
       \__     _____/
          Input

This means that (when dropout=True) we can differentiate with respect to the right pathway to get gradients (https://github.com/mdenil/dropout/blob/master/mlp.py#L260) but we can compute test error using the left pathway (https://github.com/mdenil/dropout/blob/master/mlp.py#L239). When dropout=False the right pathway isn't used at all, but the code still builds the whole graph anyway.

[saatvikshah]

Thanks for the explanation! Its definitely more comprehensible now. Whats the reason though for allowing someone using dropout to compute costs from the left pathway as well?

[mdenil]

I use the right plathway to compute test error.