stochastic optimisation (noisy gradients due to data-subsampling) for ALL models

[thangbui]

Make sure that the gradients are correct and the variance becomes smaller when the batch size gets larger.

For each model, write an example use case.