We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations

DBLP - CS Bibliography


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: A Reparameterization-Invariant Flatness Measure for Deep Neural Networks

Abstract: The performance of deep neural networks is often attributed to their automated, task-related feature construction. It remains an open question, though, why this leads to solutions with good generalization, even in cases where the number of parameters is larger than the number of samples. Back in the 90s, Hochreiter and Schmidhuber observed that flatness of the loss surface around a local minimum correlates with low generalization error. For several flatness measures, this correlation has been empirically validated. However, it has recently been shown that existing measures of flatness cannot theoretically be related to generalization due to a lack of invariance with respect to reparameterizations. We propose a natural modification of existing flatness measures that results in invariance to reparameterization.
Comments: 14 pages; accepted at Workshop "Science meets Engineering of Deep Learning", 33rd Conference on Neural Information Processing Systems (NeurIPS 2019)
Subjects: Machine Learning (cs.LG); Machine Learning (stat.ML)
Cite as: arXiv:1912.00058 [cs.LG]
  (or arXiv:1912.00058v1 [cs.LG] for this version)

Submission history

From: Linara Adilova [view email]
[v1] Fri, 29 Nov 2019 20:05:35 GMT (749kb,D)

Link back to: arXiv, form interface, contact.