Fractional moment-preserving initialization schemes for training fully-connected neural networks