Algorithm 2.

FairALM: DeepNet Classifier

1: Notations: Dual step size η, Primal step size τ

2: Input: Training Set D

3: Initializations: λ0 = 0, w0

4: for t = 0, 1, 2, …, T do

5:  Sample z ~ D

6:  Pick $v_t\in \partial \left({\widehat{e}}_{h_w}(z)+\left({\lambda }_t+\eta \right){\widehat{\mu }}_{h_w}^{s_0}(z)-\left({\lambda }_t-\eta \right){\widehat{\mu }}_{h_w}^{s_1}(z)\right)$

7:   (Primal) wt ← wt−1 – τvt

8:   (Dual) ${\lambda }_{t+1}\leftarrow {\lambda }_t+\eta \left({\widehat{\mu }}_{h_{w_t}}^{s_0}(z)-{\widehat{\mu }}_{h_{w_t}}^{s_1}(z)\right)$

9: end for

10: Output: wT