Utilizing Energy-Quality Trade-Off for Low-Cost ML-Based Compressive Sensing Reconstruction

Recent researches of compressive sensing (CS) reconstruction actively utilize machine learning (ML) algorithms for better reconstruction quality and processing latency. However, existing approaches typically disregard the energy-aware reconstruction processing while only focusing on obtaining a higher reconstruction quality at the cost of excessive energy consumption. Targeting the resource-limited edge-level systems, for the first time, this paper fairly evaluates the trade-offs between the algorithm-level performance and the energy consumption of the recent ML-based CS reconstruction schemes. Also, we present an advanced ML model to provide the attractive trade-off by sharing the trained weights for every reconstruction iteration, which is more suitable for edge-level computing platforms.