No feasible algorithm should be able to distinguish the outputs of two CSPRNGs. Are there any research papers that present results of actually testing this?
In particular, I'm interested in references that try (and hopefully fail) to distinguish them using neural networks or other machine learning techniques.
N.B.: I'm a mathematician, not a cryptographer, so – even if such references exists – I don't even know what journals to look in. My own research efforts didn’t turn up anything usable.