Seminarium "Modeling of Complex Systems"

Machine learning not only succeeds in everyday tasks like character and voice recognition problems, fingerprint identification, e-mail spam filtering, autonomously driving cars, credit card fraud detection, health care, financial modeling, and many more, but also has been already applied to quantum chemistry and material science problems. In physics, main contributions concerned the improved variational ansatz for many-body problems, wave-function reconstruction, but above all – detection of phases from synthetic or experimental data. Can this approach teach us something new about phase classification? So far it has only enabled the recovery of known phase diagrams or the location of phase transitions with qualitative agreement with more conventional approaches (but at much lower computational cost). Can we be even sure that the machine learns anything related to order parameter? Not really. These two questions should be addressed in order to fully legimitize the use of machine learning methods at least in this type of physical problems.The seminar aims to give a friendly introduction to the topic of machine learning and its use in quantum physics, and to present influence functions, being an interpretation tool developed in machine learning community. Preliminary results from studies on this method will be shown, and their potentially fruitful application on the border between machine learning and phase classification problems will be discussed.