1: \begin{abstract}

2: In this paper we address the problem of solving ill-posed inverse problems in imaging where the prior is a neural generative model.

3: %and the degradation model is log-concave.

4: Specifically we consider the decoupled case where the prior is trained once and can be reused for many different log-concave degradation models without retraining. Whereas previous \MAP-based approaches to this problem lead to highly non-convex optimization algorithms, our approach computes the joint (space-latent) \MAP\ that naturally leads to alternate optimization algorithms and to the use of a stochastic encoder to accelerate computations.

5: The resulting technique is called JPMAP because it performs Joint Posterior Maximization using an Autoencoding Prior.

6: We show theoretical and experimental evidence that the proposed objective function is quite close to bi-convex. Indeed it satisfies a weak bi-convexity property which is sufficient to guarantee that our optimization scheme converges to a stationary point.

7: %which would pave the way to show strong convergence results of our optimization scheme.

8:  Experimental results also show the higher quality of the solutions obtained by our JPMAP approach with respect to other non-convex \MAP\ approaches which more often get stuck in spurious local optima.

9: \end{abstract}

10: