Examenspresentation av Tryggve Svensson

Examiner: Giuseppe Durisi

Opponent: Thiago Puppin Romano

Abstract:
In this thesis, a comparative study is performed with respect to two CNN-type of architectures for one-shot learning, based on a relation network (RN) and a siamese network (SN), in the application of automatic modulation recognition (AMR). It is shown that the architectures for RN and SN under consideration, in some cases, are almost equal in one-shot accuracy, but that the RN more often outperforms the SN. The thesis furthermore investigates the application of the wavelet scattering transform (WST) as a way to initially represent the signals before their use in said architectures. The results demonstrate that the WST in some cases can improve the recognition accuracy in the one-shot scenario when considering the RN especially. However, the results suggest that different setups of training classes influence this improvement. As such, further research is needed to improve the way the scattering coefficients are combined with the RN, as well as to explore additional ways to enhance the overall one-shot accuracy.

Welcome!
Tryggve and Giuseppe