Algorithm 1 Patch learning (PL)
Input: Training data $D={\{({\mathbf{x}}_i,{\mathbf{y}}_i)\}}_{i=1}^N$, ML matrix W2, Lagrange multiplier of ADMM ρ and U, learning rate η1, and penalty parameter α.
Output: PL matrix W1 ∈ ℝD×L with sparse groups of rows.
1:	for ℓ = 1, … , L do
2:	${\mathbf{w}}_{1\ell }^{(0)}=\frac{1}{D}{\mathbf{1}}_D,{\mathbf{v}}^{(0)}=\frac{1}{D}{\mathbf{1}}_D,a^{(0)}=1,t=0$; // Initialization
3:	while not convergence do
4:	${\mathbf{z}}^{(t)}={\mathbf{v}}^{(t)}-{\eta }_1(\nabla L({\mathbf{w}}_{1\ell }^{(t)},D)+{\mathbf{u}}_{\ell }^{(t)}+\rho ({\mathbf{w}}_{1\ell }^{(t)}-{\mathbf{w}}_{2\ell }^{(t)}))$;
5:	for p = 1, … , 49 do
6:	${\mathbf{w}}_{1\ell }^{p(t+1)}=I\left(\Vert {\mathbf{z}}^{p(t)}{\Vert }_2>\alpha \right)\left(1-\frac{\alpha }{\Vert {\mathbf{z}}^{p(t)}{\Vert }_2}\right){\mathbf{z}}^{p(t)}$;
	// ${\mathbf{w}}_{1\ell }^p$ is the p-th patch within the ℓ-th column of W1
7:	end for
8:	$a^{(t+1)}=\frac{2}{t+1}$;
9:	${\mathbf{v}}^{(t+1)}={\mathbf{w}}_{1\ell }^{(t+1)}+(\frac{1-a^{(t)}}{a^{(t)}}{a}^{(t+1)})({\mathbf{w}}_{1\ell }^{(t+1)}-{\mathbf{w}}_{1\ell }^{(t)})$;
10:	t = t + 1;
11:	end while
12:	end for