Table 3. Pseudo-code of Algorithm 2.

Algorithm 2 Parameter learning for LEAE

Input: The training set ${\left\{X^{(i)}\right\}}_{i=1}^m$.

Output: the parameter of embedding mapping W, b, c;

Randomly shuffle the training set. Randomly initialized W, b, c

For each epoch

For each batch instances ${\left\{X^{(i)}\right\}}_{i=1}^n$

Select the discriminative neighbors set of X(i);

Calculate the activations for the hidden layer and output layer via a feedforward pass;

Compute the partial derivatives in regard to the input as Eqs 11, 12 and 13

Compute: $\Delta W={\displaystyle {\sum }_{n,i,j}\frac{\partial J}{\partial W_{ij}}}$;$\Delta b={\displaystyle {\sum }_{n,i}\frac{\partial J}{\partial b_i}}$;$\Delta c={\displaystyle {\sum }_{n,j}\frac{\partial J}{\partial c_j}}$

Update: $W=W-\alpha \left(\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$n$}\right.\Delta W\right)$;$b=b-\alpha \left(\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$n$}\right.\Delta b\right)$;$c=c-\alpha \left(\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$n$}\right.\Delta c\right)$

End for

End for