Regensburg 2016 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

BP: Fachverband Biologische Physik

BP 53: Posters - Statistical Physics of Biological Systems

BP 53.2: Poster

Mittwoch, 9. März 2016, 17:00–19:00, Poster C

Stochastic thermodynamics of learning in neural networks — •Sebastian Goldt and Udo Seifert — II. Institut für Theoretische Physik, Universität Stuttgart, 70550 Stuttgart, Germany

Over the past decade, stochastic thermodynamics has emerged as a powerful framework to understand the role of information in physical systems and the thermodynamic costs of manipulating it. A particularly intriguing application of these ideas is biology: every organism first gathers information about its noisy environment and then builds models from that data, at the expense of energy dissipation. Here, we focus on the second part of this process: learning.

Biologically, learning is implemented in neural networks where neurons receive and send signals from and to many other neurons via synapses. The strength of these synapses determines whether an incoming signal will make the neuron trigger an action potential, the electric pulse that is the basic token of communication in neural systems. The adaptation of synapses is the physiological mechanism for memory formation, e.g. in Hebbian learning.

Here, we use stochastic thermodynamics to analyse the learning of a classification rule by a feedforward neural network, whose synapses we endow with stochastic dynamics. Starting from the total entropy production of the network, we identify the rate of learning in a thermodynamically consistent way and introduce a measure for the thermodynamic efficiency of learning, which we compute for different learning algorithms.