Journal article

Gaussian mean field regularizes by limiting learned information

  • Kunze, Julius Computer Science, University College London, London WC1E 6BT, UK
  • Kirsch, Louis Computer Science, University College London, London WC1E 6BT, UK - Istituto Dalle Molle di studi sull'intelligenza artificiale (IDSIA), Facoltà di scienze informatiche, Università della Svizzera italiana, Svizzera
  • Ritter, Hippolyt Computer Science, University College London, London WC1E 6BT, UK
  • Barber, David Computer Science, University College London, London WC1E 6BT, UK - Alan Turing Institute, London NW1 2DB, UK
Show more…
    03.08.2019
Published in:
  • Entropy. - 2019, vol. 21, no. 8, p. 758
English Variational inference with a factorized Gaussian posterior estimate is a widely-used approach for learning parameters and hidden variables. Empirically, a regularizing effect can be observed that is poorly understood. In this work, we show how mean field inference improves generalization by limiting mutual information between learned parameters and the data through noise. We quantify a maximum capacity when the posterior variance is either fixed or learned and connect it to generalization error, even when the KL-divergence in the objective is scaled by a constant. Our experiments suggest that bounding information between parameters and data effectively regularizes neural networks on both supervised and unsupervised tasks.
Language
  • English
Classification
Computer science
License
License undefined
Identifiers
Persistent URL
https://susi.usi.ch/usi/documents/318902
Statistics

Document views: 18 File downloads:
  • Kunze_E_2019.pdf: 5