Translations:Batch Normalization Accelerating Deep Network Training/5/en

batch normalization addresses this by normalizing the inputs to each layer using statistics computed over the current mini-batch. This ensures that each layer receives inputs with a stable mean and variance, regardless of changes in preceding layers. The technique is applied as a differentiable transformation inserted into the network architecture, making it compatible with standard backpropagation and stochastic gradient descent.