1: \begin{abstract}

2: We propose a deep-learning approach based on 

3: generative adversarial networks (GANs) to reduce noise in weak lensing mass maps under realistic conditions. 

4: We apply image-to-image translation using conditional GANs to the mass map obtained from the first-year data of Subaru Hyper Suprime-Cam (HSC) survey. 

5: We train the conditional GANs by using 25000 mock HSC catalogues that directly incorporate a variety of observational effects. 

6: We study the non-Gaussian information in denoised maps using

7: one-point probability distribution functions (PDFs) and also perform matching analysis for positive peaks and massive clusters.

8: An ensemble learning technique with our GANs is successfully applied to reproduce the PDFs of the lensing convergence. About $60\%$ of the peaks in the denoised maps with height greater than $5\sigma$ have counterparts of massive clusters within a separation of 6 arcmin.

9: We show that PDFs in the denoised maps are not compromised by 

10: details of multiplicative biases and photometric redshift distributions, 

11: nor by shape measurement errors, 

12: and that the PDFs show stronger cosmological dependence compared to the noisy counterpart.

13: We apply our denoising method to a part of the first-year HSC data to show that the observed mass distribution is statistically consistent with the prediction from the standard $\Lambda$CDM model.

14: \end{abstract}

15: