Sharp Minima Can Generalize For Deep Nets
, , , and .
(2017)cite arxiv:1703.04933Comment: 8.5 pages of main content, 2.5 of bibliography and 1 page of appendix.

Despite their overwhelming capacity to overfit, deep learning architectures tend to generalize relatively well to unseen data, allowing them to be deployed in practice. However, explaining why this is the case is still an open area of research. One standing hypothesis that is gaining popularity, e.g. Hochreiter & Schmidhuber (1997); Keskar et al. (2017), is that the flatness of minima of the loss function found by stochastic gradient based methods results in good generalization. This paper argues that most notions of flatness are problematic for deep models and can not be directly applied to explain generalization. Specifically, when focusing on deep networks with rectifier units, we can exploit the particular geometry of parameter space induced by the inherent symmetries that these architectures exhibit to build equivalent models corresponding to arbitrarily sharper minima. Furthermore, if we allow to reparametrize a function, the geometry of its parameters can change drastically without affecting its generalization properties.
  • @jk_itwm
  • @dblp

rating distribution
average user rating5.0 out of 5.0 based on 1 review
  • userPicture
    a month ago (last updated a month ago)
Please log in to take part in the discussion (add own reviews or comments).