Table 3: Summary of the regularization results from neural networks with different training sets and weight decay term $\gamma$ . Increasing $\gamma$ results in smaller weights of the network and thus a more regularized solution. However, too large a value of $\gamma$ will again result in larger deviations, i.e. there is a trade-off in setting this parameter. The metallicity deviations for the standard stars are given in terms of the median values of the difference (computed value - literature value). It can be seen that training on noise-free data and validating on noisy data systematically underestimated metallicities. The results demonstrate that noise in the network inputs can help improve the regularization.
Noise in training set $\gamma$ [Fe/H] offset

no 0.0001 -0.16

no 0.001 -0.11

no 0.01 -0.07

yes 0.0001 0.18

yes 0.001 0.02

yes 0.01 0.05

**Table 3:** Summary of the regularization results from neural networks with different training sets and weight decay term $\gamma$ . Increasing $\gamma$ results in smaller weights of the network and thus a more regularized solution. However, too large a value of $\gamma$ will again result in larger deviations, i.e. there is a trade-off in setting this parameter. The metallicity deviations for the standard stars are given in terms of the median values of the difference (computed value - literature value). It can be seen that training on noise-free data and validating on noisy data systematically underestimated metallicities. The results demonstrate that noise in the network inputs can help improve the regularization.
Noise in training set	$\gamma$	[Fe/H] offset
no	0.0001	-0.16
no	0.001	-0.11
no	0.01	-0.07
yes	0.0001	0.18
yes	0.001	0.02
yes	0.01	0.05

Source LaTeX | All tables | In the text

	1 Introduction
	2 Observations and reductions
	3 Grid of observed spectra
	4 Grid of synthetic spectra
	5 Artificial neural networks
	6 Validation of the network model
	7 Concluding remarks
	References