Deriving the Glorot initialization scheme The pre-2010 phase of deep learning research made extensive use of model pre-training, since training deep models was thought to be very hard. This changed in 2010 due in part to a paper by Xavier Glorot and Yoshua Bengio who showed that deep models can be trained by just ensuring good initializations. The key insight by Xavier was that a layer in a deep network should ensure that data passed through it maintains the same variance, since if it does not, deeper networks will have a multiplicative effect and change the variance even more. Derive the Glorot initialization scheme for a relu layer using this principle. Use the constraint on both forward and backward passes.