Berkeley Statistics and Machine Learning Forum
Denoising
Deconvolution
Inpainting
Where does my prior come from ?
Wavelet Sparsity
Total Variation
Gaussian
Example in MRI: Deep ADMM-Net for Compressive Sensing MRI
Credit: Lustig et al. 2008
This is great, but can we do better?
Recurrent Inference Machines for Solving Inverse Problems
Putzky & Welling, 2017
Why not write this iteration as:
To match the RNN framework, an additional variable s is introduced to store an optimization memory state.
We trained three models on these tasks: (1) a Recurrent Inference Machine (RIM) as described in 2, (2) a gradient-descent network (GDN) which does not use the current estimate as an input (compare Andrychowicz et al. [15]), and (3) a feed-forward network (FFN) which uses the same inputs as the RIM but where we replaced the GRU unit with a ReLu layer in order to remove hidden state dependence.
Super-resolution example with factor 3
Recurrent Inference Machines
for Accelerated MRI Reconstruction
Lønning et al. 2018
All models in this paper were trained on acceleration factors that were randomly sampled from the uniform distribution U (1.5, 5.2). Sub-sampling patterns were then generated using a Gaussian distribution.
Data-Driven Reconstruction of Gravitationally Lensed Galaxies using Recurrent Inference Machines
Morningstar et al. 2019