Taking advantage of noisy spectra through machine learning

Traditionally, noise is considered a limiting factor in (experimental) spectra. Here we demonstrate with two rather different examples how noise can be turned into an asset with the help of Machine Learning. In the first example we purify successfully noisy photo electron spectra generated with intense XFEL pulses. The single shot spectra contain in the combination of non-linear light-matter coupling and broad-band noise much more information about the target than a clean spectrum. In the second example we construct a system of networks with different levels of noise trained to recognize periodic signals. The system identifies correctly the periodicity as well as the noise level in a spectrum, as demonstrated with intrinsically noisy high harmonic spectra.