Xavier Initialization

• $\mathrm{a}=\sqrt{\frac{6}{\left(\mathrm{d}{}_{\mathrm{in}}^{}+{\mathrm{d}}_{\mathrm{out}}\right)}}$
• Random values drawn uniformly from $[-a,a]$
• For Batch Normalization Layers, $\gamma =1$ and $\beta =0$
• For Tanh based activating neural nets