nce than the standard or classic algorithms for pattern discovery

gnition [Waldrop, 2019].

nately, several new ideas have been motivated such as the Lasso

twork [Ross and Lage, 2017; Mohammadi, 2020] and the sparse

ral network [Scardapane, et al., 2017; Qiao, et al., 2020]. The

ural network can increase the interpretation capability using a L1

pplied to the input gradient shown below, where x stands for

stands for labels, Hሺݕ, ݕොሻ stands for the cross-entropy,

ߠ|ݔ, ݕሻൌܪሺݕ, ݕොሻ൅ߣఏ‖ߠ‖ଵ൅ߣ׏ ቛ׏௫ܪቀ^ଵ

௄^{, ݕොቁቛ}

ଵ

(9.1)

onventional approach for generating a sparse neural network as

deep learning model for increasing the interpretation capability

y a regularisation constraint to the model parameters (weights).

her alternative, the group sparse regularisation deep learning

aims to group neurons or regularise neurons to achieve a better

erpretation capability [Scardapane, et al., 2017; Lin and Runger,

ao, et al., 2020]. All of these new efforts aim to explore the

ation capability of a neural network model or a deep network

is expected that these new developments may play an increasing

ological/medical pattern analysis projects soon. To increase the

ce of a neural network model or a deep neural network model, it

n well recognised that few neurons can promote a better

ce power of a network [Lechner, et al., 2020]. But how to reduce

l complexity while maintaining the model accuracy is still an

ject for research. The promising ideas of new research suggest

bination between artificial intelligence with deep learning to

more interpretable machine learning models [Heaven, 2019].

misation with evolutionary computation

etic algorithm [Holland, 1992], the genetic programming

m [Koza, 2010] and the evolutionary algorithms [Back and

, 1993] belong to another important area of machine learning