ML Model Compression Techniques for Interference Cancellation

Sammanfattning

Passive Intermodulation (PIM) is a phenomenon that occurs in passive components of wireless communication systems, leading to unwanted signal distortion and interference. Although PIM has traditionally been canceled using polynomial models, machine learning approaches have recently shown improved cancellation performance. However, low latency is essential for inference on the resource-limited edge devices used for real-time applications like PIM Cancellation. This work examines machine learning model compression techniques for this interference cancellation, reducing model complexity while avoiding a reduction in cancellation performance. First, we pinpoint appropriate strategies for the specified use case, including conventional and experimental methods such as Binary Neural Networks. We then apply these techniques to our Neural Networks and specific interference cancellation test cases to analyze their effectiveness. We show that some approaches can be used with almost no adverse effect, while others need further consideration regarding the latency performance trade-off. We also show that combining methods is a viable strategy to increase compression. We conclude with a discussion on implementing the discovered compression techniques as a relevant problem in the field of model compression and propose a non-destructive soft neuron pruning algorithm. Although there are numerous obstacles to overcome when implementing ML-based PIMC on live radio networks, the techniques presented in this thesis broaden our understanding and bring us one step closer to making it a reality.