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

Download:

Current browse context:

cs.LG

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

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

Computer Science > Machine Learning

Title: Probing the State of the Art: A Critical Look at Visual Representation Evaluation

Abstract: Self-supervised research improved greatly over the past half decade, with much of the growth being driven by objectives that are hard to quantitatively compare. These techniques include colorization, cyclical consistency, and noise-contrastive estimation from image patches. Consequently, the field has settled on a handful of measurements that depend on linear probes to adjudicate which approaches are the best. Our first contribution is to show that this test is insufficient and that models which perform poorly (strongly) on linear classification can perform strongly (weakly) on more involved tasks like temporal activity localization. Our second contribution is to analyze the capabilities of five different representations. And our third contribution is a much needed new dataset for temporal activity localization.
Comments: erd59xH@Rqt!XCMsCmnz
Subjects: Machine Learning (cs.LG); Computer Vision and Pattern Recognition (cs.CV); Machine Learning (stat.ML)
Cite as: arXiv:1912.00215 [cs.LG]
  (or arXiv:1912.00215v2 [cs.LG] for this version)

Submission history

From: Cinjon Resnick [view email]
[v1] Sat, 30 Nov 2019 15:05:28 GMT (1468kb,D)
[v2] Thu, 12 Aug 2021 17:54:53 GMT (1468kb,D)

Link back to: arXiv, form interface, contact.