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

Download:

Current browse context:

cs.SD

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 ScienceWISE logo

Computer Science > Sound

Title: Redundancy Reduction Twins Network: A Training framework for Multi-output Emotion Regression

Abstract: In this paper, we propose the Redundancy Reduction Twins Network (RRTN), a redundancy reduction training framework that minimizes redundancy by measuring the cross-correlation matrix between the outputs of the same network fed with distorted versions of a sample and bringing it as close to the identity matrix as possible. RRTN also applies a new loss function, the Barlow Twins loss function, to help maximize the similarity of representations obtained from different distorted versions of a sample. However, as the distribution of losses can cause performance fluctuations in the network, we also propose the use of a Restrained Uncertainty Weight Loss (RUWL) or joint training to identify the best weights for the loss function. Our best approach on CNN14 with the proposed methodology obtains a CCC over emotion regression of 0.678 on the ExVo Multi-task dev set, a 4.8% increase over a vanilla CNN 14 CCC of 0.647, which achieves a significant difference at the 95% confidence interval (2-tailed).
Comments: 5 pages, accepted by ICML Exvo workshop
Subjects: Sound (cs.SD); Audio and Speech Processing (eess.AS)
Cite as: arXiv:2206.09142 [cs.SD]
  (or arXiv:2206.09142v2 [cs.SD] for this version)

Submission history

From: Xin Jing [view email]
[v1] Sat, 18 Jun 2022 07:56:02 GMT (449kb,D)
[v2] Tue, 28 Jun 2022 09:58:25 GMT (449kb,D)

Link back to: arXiv, form interface, contact.