We consider the problem of recovering a structured (e.g., sparse) signal from noisy high-dimensional linear measurements.
In the recent past, the dominant approach to signal recovery was algorithmic.
But, nowadays, algorithms are increasingly being replaced by deep neural networks (DNNs) that can learn optimal inference strategies directly from the data.
In this talk, we propose novel feedforward DNN architectures that are inspired by the Onsager correction mechanism in approximate message passing (AMP) algorithms.
Preliminary results show that the proposed networks are highly efficient for our signal-recovery problem. 
Moreover, the optimal DNN parameters are directly interpretable and can be deduced in a straightforward way from the signal statistics, when known.