Abstract
The task was to recognize anomalies in RF recordings contain I/Q signal presentation. The signal transformed to sequence of vector time-series of Power Spectral Density (PSD) vectors.
TSG have a program implementing a LSTM-LSTM GAN algorithm that recognizes anomalies with 85-90% of success. The main goal of the project was to improve the disadvantages this algorithm has, such as: improve running time, make it fully unsupervised and try to improve the accuracy.
Challenges
1- Working with an existed code. It took me a long time to go over the code and be able to make changes in it.
2- The project involved a lot of research and new techniques. In other words, it means to implement deep learning methods without having any references or examples. In addition, we did not know what the outcome will be and what is considered to be a good result.
Achievements (according to KPIs)
1- Able to implement SRU-GAN technique and gain the same accuracy results with an improvement of 3.2x speed in compare to the LSTM-GAN the company had.
2- Use Wasserstein loss function instead of the regular KL loss used in the GAN.
3- Combine between 1 and 2.
Further development
The WGAN-SRU we got still does not give us the high accuracy results in comparison to the GAN-SRU. In addition, the algorithm is not fully unsupervised yet. That will require some further work.
Supervisor Feedback
This Data Science Fellows project is part of TSG’s much wider and complex activities. Algorithms development, implementation and improvement is the essential part of this activity.
The task given to Rotem was to improve the performance of LSTM-LSTM GAN used in TSG with SRU-SRU GAN. SRU method is reported in literature as working significantly faster than LSTM. This work was performed by Rotem perfectly. The results obtained in her work will be implemented in TSG projects. As a result of this student activity we recommended to recruit Rotem to be a part of the TSG team.
