Receive a weekly summary and discussion of the top papers of the week by leading researchers in the field.

In Optics letters

Deep learning (DL) has been used to successfully solve numerous problems and challenges in a wide range of fields. The architecture of DL is complex and treated as a black box, making it difficult to understand the principles behind it. Here, we visualize the process of compensating for time mismatches for a two-channel photonic analog-to-digital converter (PADC) by a convolutional recurrent autoencoder (CRAE) with excellent generalizability and robustness. Besides, we explore the effects of different modules of the CRAE on the generalizability. Based on the analysis of the above two operations, we simplify the CRAE and then apply it to a four-channel PADC, which is a more complex channel-interleaved system. Consequently, for both PADC systems, the performance of the simplified CRAE is as good as that of the original CRAE. Moreover, for the two-channel PADC, after simplification, the frame rate of the CRAE is increased from 460 frames/second to 975 frames/second, 20,000 points in each frame. For the four-channel PADC, the spur-free dynamic range is enhanced to 24.6 dBc from 5.2 dBc.

Zou Xiuting, Xu Shaofu, Zou Weiwen