The aim of this thesis has been to look at the possibility of using a convolutional neural network to attenuate seismic interference (SI) noise in marine seismic data. Modern SI-denoising algorithms employed by the industry are efficient and normally yield very good results. However, they are often time consuming. Using neural networks to do real time denoising could significantly improve denoising efficiency, and thereby save time and money for processing companies. The dataset used in this thesis consisted of two lines of marine seismic field data, where one line was denoised marine seismic data and the second line contained "pure" seismic interference noise recorded in 2015 in the North Sea. These were combined to create datasets needed to train the convolutional neural networks. Four different models were applied in this thesis: Classification CNN, Autoencoder, No Downscaling CNN and U-NET. The Classification CNN was implemented asa proof of concept to test whether the network could differentiate between noise-contaminated data and clean data, which yielded good results. The autoencoder gave poor denoising results, showing significant loss of geological signal. Both the No Downscaling CNN and the U-NET gave good results, where U-NET performed best. It performed good denoising, leaving only a very small residual in special cases of conflicting dip. The network required approximately 0.02s for denoising a shot gather, proving that real time denoising of marine seismic data is possible with the use of convolutional neural networks.