Fast Wiener filtering of CMB maps with Neural Networks

Abstract: We show how a neural network can be trained to Wiener filter masked CMB maps to high accuracy. We propose an innovative neural network architecture, the WienerNet, which guarantees linearity in the data map. Our method does not require Wiener filtered training data, but rather learns Wiener filtering from tailored loss functions which are mathematically guaranteed to be minimized by the exact solution. Once trained, the neural network Wiener filter is extremely fast, about a factor of 1000 faster than the standard conjugate gradient method. Wiener filtering is the computational bottleneck in many optimal CMB analyses, including power spectrum estimation, lensing and non-Gaussianities, and our method could potentially be used to speed them up by orders of magnitude with minimal loss of optimality. The method should also be useful to analyze other statistical fields in cosmology.