Hybridizing Evolutionary Computation and Deep Neural Networks: An Approach to Handwriting Recognition Using Committees and Transfer Learning

Algorithm 1

Definition of the grammar in Backus-Naur Form for the MNIST dataset. Source: Baldominos et al. [3].

<dnn> ⩴ <input>  <conv_lys>  <dense_lys>  <opt_setup>

<input> ⩴ <batch_size>

<batch_size> ⩴ 25 | 50 | 100 | 150

<conv_lys> ⩴ <conv> | <conv>  <conv> | <conv>  <conv>  <conv>

<conv> ⩴ <n_kernels>  <k_size>  <act_fn>  <pooling>

<n_kernels> ⩴ 8 | 16 | 32 | 64 | 128 | 256

<k_size> ⩴ 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9

<pooling> ⩴ null | <p_size>

<p_size> ⩴ 2 | 3 | 4 | 5 | 6

<dense_lys> ⩴ <dense> | <dense>  <dense> | <dense>  <dense>  <dense>

<dense> ⩴ <d_type>  <n_units>  <act_fn>  <reg_fn>  <dropout_r>

<d_type> ⩴rnn | lstm | gru | feedforward

<n_units> ⩴ 32 | 64 | 128 | 256 | 512 | 1024

<act_fn> ⩴relu | linear

<reg_fn> ⩴ null | l1 | l2 | l1l2

<dropout_r> ⩴ 0 | 0.5

<opt_setup> ⩴ <opt_type>  <learn_rate>  <batch_size>

<opt_type> ⩴sgd | nesterov | momentum | adagrad | adamax | adam | adadelta | rmsprop

<learn_rate> ⩴ 5E-1 | 1E-1 | 5E-2 | 1E-2 | 5E-3 | 1E-3