diff --git a/README.md b/README.md index 51bd8ad..78cd7f7 100644 --- a/README.md +++ b/README.md @@ -41,8 +41,7 @@ training time. To achieve this, we rewrite the backward pass of BN in terms of its output `y`, which is in turn reconstructed from `z` by inverting the activation function. -The parametrization for the scaling factor of BN changed compared to standard BN, in order to ensure an invertible transformation. Specifically, the scaling factor becomes -. +The parametrization for the scaling factor of BN changed compared to standard BN, in order to ensure an invertible transformation. Specifically, the scaling factor becomes $\hat{\gamma} = \left | \gamma \right | + \epsilon$ . ## Requirements