Scaling Auto-Unrolled Proximal Gradient Descent: AutoML for Physical-Layer Optimization
By utilizing AutoGluon to automate hyperparameter tuning for unrolled Proximal Gradient Descent architectures, engineers can achieve 98.8% of the spectral efficiency of a 200-iteration solver with only 5 unrolled layers, significantly reducing inference latency at the cost of requiring domain-specific gradient normalization.