Algorithm 1 Neuron agglomerative clustering.

1:For each layer l in the network: 2:Let $ne{u}_i^l=[{\mathbf{W}}_{i,1}^l,{\mathbf{W}}_{i,2}^l,{\mathbf{W}}_{i,3}^l,\cdots ,{\mathbf{W}}_{i,n_{l-1}}^l,{\mathbf{b}}_i^l]$; 3:Construct neuron set $S_l=\{ne{u}_1^l,ne{u}_2^l,ne{u}_3^l,\cdots ,ne{u}_{n_l}^l\}$; 4:Cluster the neurons in set $S_l$ into $k_l$ groups using agglomerative clustering; 5:Construct a set of cluster centroids $P_l=\{p_1^l,p_2^l,p_3^l,\cdots ,p_{k_l}^l\}$; 6:Agglomerate neurons of the same cluster into its cluster centroid; 7:Remember the agglomerating list $R_l=[r_1^l,r_2^l,r_3^l,\cdots ,r_{n_l}^l]$; 8:Calculate ${\tilde{\mathbf{W}}}_{i,p}^{l+1}={\sum }_{j=1}^{n_l}\mathbb{I}(r_j^l=p){\mathbf{W}}_{i,j}^{l+1}$, $i=1,\cdots ,n_{l+1}$, $p=1,\cdots ,k_l$, where $\mathbb{I}(·)$ is the indication function; 9:Agglomerate connections of layer $l+1$ into ${\tilde{\mathbf{W}}}_{i,p}^{l+1}$. Bias remains unchanged.