Dynamic Mechanisms in Neural Networks

• Oscillations
• Statistics of correlations
• Dimensionality
• Chaos

Mechanisms of Neural Information Processing

• Biological and artificial neural networks
• Generalization properties
• Field theory of feedforward and recurrent networks

We employ methods of statistical physics to understand information processing in biological and artificial neuronal networks (physics of AI). Using methods from equilibrium and non-equilibrium and statistical field theory, we study learning in the setting of Bayesian inference to obtain insights into the expressibility and generalization properties of neuronal networks. This mathematical language applies to both, biological and artificial network architectures and allows us to distill common underlying mechanisms and to quantify differences. Collective dynamical properties, such as chaotic dynamics and the closeness to critical points, for example decisively shape learning and generalization in recurrent biological networks and artificial feed forward ones.

Methods Transfer and Adaptation from Theoretical Physics

• Statistical field theory
• Theory of disordered systems
• Methods from Statisitcal Physics (Fokker Planck theory, Edgeworth expansion, etc.)