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

DY: Fachverband Dynamik und Statistische Physik

DY 1: Focus Session: Artificial Intelligence in Condensed Matter Physics I (joint session TT/DY)

DY 1.2: Hauptvortrag

Montag, 18. März 2024, 10:00–10:30, H 0104

Communicability as a criterion for interpretable representations — •Renato Renner — ETH Zürich, Zürich, Switzerland

We propose an autoencoder architecture that can generate representations of data from physical experiments which are operationally meaningful and thus interpretable. The architecture is based on the paradigm of ``communicability''. Roughly, the idea is that the encoder orders the data into several parts that may be communicated separately to agents, whose task is to answer different questions about the data. The encoding is then optimised so that this communication is minimised, i.e., each agent receives precisely the information that is relevant to its task. Using some toy examples, including ones from quantum state tomography, we show that this approach leads to a separation of parameters, which can be regarded as a step towards interpretability.

Keywords: Neural networks; interpretable representations; information theory