Try replacing Batch Normalization with SELU, and make the necessary adjustments to ensure the network self$-$normalizes $($i$.$e$.,$ standardize the input features, use LeCun normal initialization, make sure the DNN contains only a sequence of dense layers$).$