Budapest Artificial Intelligence Meetup - November

Presentations: http://study.inf.elte.hu/aimonday/

• Márk Jelasity, DSc Habil (University of Szeged, Department of Computer Algorithms and Artificial Intelligence):

Adversarial learning: the effect of dimensionality and regularization on linear models

In adversarial learning one is concerned with the recently discovered
phenomenon that most machine learning models are extremely sensitive to adversarial examples. That is, for most correctly classified examples
there exist extremely similar examples that are misclassified. This
phenomenon can be observed even in the case of the simplest linear models. We examine such models and attempt to understand the geometric reasons of the existence of adversarial examples. We argue that dimensionality and regularization play a very important role.

• Pataki Bálint Ármin, PhD student (ELTE, Department of Physics of Complex Systems):

More accurate cosmological parameter constraints via deep learning

Cosmological models can describe our universe, although they have some parameters that cannot be directly obtained. We must make observations - in this case sky surveys - to make constraints for these parameters.

Luckily we can also make simulations with given cosmological parameter sets.

Creation of a model that can retrieve the parameters from the results of the simulations can help us to retrieve the parameters for the real simulation, the Universe. As in almost all of the image based topics, deep learning is involved here too. But surprisingly one of the kernels has physical meaning, which lead us to a non black-box model with similar performance.