Entanglement is the best known (and strongest) type of quantum correlations, but there exist also weaker forms, such as the quantum discord. For mixed states, the different correlations differ from each other. A state is entanglement when it is not possible to create it using only classical means (local operations and classical communication), but discordance is defined via the amount of information about one subsystem that can be gained by the measurement of another. The set of discordant states is larger than the set of entangled states, and any entangled state is discordant.
We used advanced machine learning tools in order to study quantum correlations from the perspective of mixed density matrices. The algorithm was trained to distinguish entangled states from separable (not entangled) states on varied sets of mixed and pure states. Later this algorithm was used in order to distinguish between separable and entangled states. We compared the obtained sets with the known true distinctions between entanglement and separability, as well as other markers of correlations (both classical and quantum). We found, quite surprisingly, that the output of the algorithm best corresponds to the set of discordant states. This means that from a machine learning perspective, discordant states are much easier to classify as quantum-correlated, even though the training process involved only entangled states.
