Learning Complexity: How Neural Networks Discover Structure

Karoline Wiesner

How do neural networks learn? A useful way to approach this question is through the idea of compression: learning can be understood as the process of finding a compact representation of data while retaining the information that matters. In this talk, I will explore structure discovery from the perspective of model complexity, where the model itself is progressively enlarged as structure is discovered. Our recent work shows that this process is closely connected to the statistical physics of phase transitions. As a model discovers structure, it passes through a hierarchy of phase transitions, each corresponding to the learning of particular features.

Weitere Infos zur Veranstaltung