This dataset consists of radio signals modulated using different modulation techniques and different signal to noise ratios. The task was to try different neural network architectures to classify the signals based on the modulation technique that is used. The performance of the networks was compared accross all signal to noise ratios. The best average accuracy on the test data was 62% which is high compared to the results achieved by the paper Deep learning for modulation recognition which worked with the same dataset.
Three different architectures where tried a plain fully connected network, a plain CNN and an inception CNN. The highest accuracy was obtained using the inception approach.The repo contains three jupyter notebooks each trying a different architecture.Each notebook has a cell at its begining with a bash script to download the dataset.Keras was used as the deep learning library to build the networks.The report.pdf contains detailed explanation of the approach and results.ProblemStatment.pdf has detailed explanation of the task as well as links to papers with more indepth explanation of the dataset.