Dresden 2017 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

DY: Fachverband Dynamik und Statistische Physik

DY 2: Stochastic thermodynamics and information processing

DY 2.4: Talk

Monday, March 20, 2017, 10:30–10:45, HÜL 186

Stochastic Thermodynamics of Learning — •Sebastian Goldt and Udo Seifert — II. Institut für Theoretische Physik, Universität Stuttgart, 70550 Stuttgart

Virtually every organism gathers information about its noisy environment and builds models from that data, mostly using neural networks. Here, we use stochastic thermodynamics to analyse the efficiency of neural networks in two learning scenarios. We show that the total entropy production of the network bounds the information that the network can infer from data or learn from a teacher [1]. We introduce a learning efficiency η≤1 and discuss the conditions for optimal learning. Finally, we analyse the efficiency of the Hebbian, Perceptron and AdaTron learning algorithms, well-known from machine learning and statistical physics.
[1] S. Goldt and U. Seifert, Stochastic Thermodynamics of Learning. PRL, in press; arxiv:1611.09428