Announcement: Berkeley workgroup on Learning, Information Theory, and Nonequilibrium Thermodynamics

maxNext meeting: Friday 3/7 at 3:30pm in 560 Evans.
We’ll be going over the recent Spinney-Ford and Jarzynski review papers.

(Inaugural meeting: Feb 21, 2014, 3.30pm, and every 2 weeks thereafter)
Location: Redwood Center Seminar Room, 560 Evans Hall, UC Berkeley

For email notices signup to the LINEQ mailing list..


Natural thematic overlap exists between statistical mechanics and thermodynamics on the one hand, and statistical machine learning and inference on the other. Are these relationships just analogies, or can we study and build systems which have adaptive models of their world and in doing so save energy?

This question is pertinent to (1) understanding the self-organisation of biophysical systems, and (2) laying the foundations of adaptive
energy-saving devices, ie: learning Maxwell’s demons. Thus answers to these questions would have much scientific and technological value. The purpose of our group is to (1) understand these issues, and (2) work towards solving them.

The meetings will be an evolving mixture of seminar, discussion and journal club format. We will start by studying the battery of results on Fluctuation Theorems, up to the Sagawa-Ueda results, and attempt to elucidate their relationships to sensory and sensory-motor unsupervised learning theory.