Hi Varun. In inference time, the layers output is calculated using the formulas stated in point “3. Scale and shift in order to obtain the output of the layer” but, instead of using the batch mean and variance to scale the input, you use a mean and variance estimated from the ones calculated during training.