Training Binary Layers by Self-Shrinking of Sigmoid Slope: Application to Fast MRI Acquisition

Filippo Martinini, Andriy Enttsel, Alex Marchioni, Mauro Mangia, Fabio Pareschi, Riccardo Rovatti, Gianluca Setti

Abstract: Deep Neural Networks (DNN) have become popular and widespread because they combine computational power and flexibility, but they may present critical hyper-parameters that need to be tuned before the model can be trained. Recently, the use of trainable binary masks in the field of Magnetic Resonance Imaging (MRI) acquisition brought new state-of-the-art results, but with the disadvantage of introducing a bulky hyper-parameter, which tuning is usually time-consuming. We present a novel callback-based method that is applied during training and turns the tuning problem into a triviality, also bringing non-negligible performance improvements. We test our method on the fastMRI dataset.