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 ScienceWISE logo

Computer Science > Information Theory

Title: DECONET: an Unfolding Network for Analysis-based Compressed Sensing with Generalization Error Estimates

Abstract: We present a new deep unfolding network for analysis-sparsity-based Compressed Sensing. The proposed network coined Decoding Network (DECONET) jointly learns a decoder that reconstructs vectors from their incomplete, noisy measurements and a redundant sparsifying analysis operator, which is shared across the layers of DECONET. Moreover, we formulate the hypothesis class of DECONET and estimate its associated Rademacher complexity. Then, we use this estimate to deliver meaningful upper bounds for the generalization error of DECONET. Finally, the validity of our theoretical results is assessed and comparisons to state-of-the-art unfolding networks are made, on both synthetic and real-world datasets. Experimental results indicate that our proposed network outperforms the baselines, consistently for all datasets, and its behaviour complies with our theoretical findings.
Comments: submitted to journal
Subjects: Information Theory (cs.IT); Information Retrieval (cs.IR); Machine Learning (cs.LG)
Cite as: arXiv:2205.07050 [cs.IT]
  (or arXiv:2205.07050v5 [cs.IT] for this version)

Submission history

From: Vicky Kouni Ms [view email]
[v1] Sat, 14 May 2022 12:50:48 GMT (337kb,D)
[v2] Sat, 21 May 2022 14:50:15 GMT (336kb,D)
[v3] Sat, 11 Jun 2022 18:33:23 GMT (331kb,D)
[v4] Wed, 10 Aug 2022 20:30:39 GMT (4825kb,D)
[v5] Sat, 13 Aug 2022 14:37:08 GMT (4825kb,D)

Link back to: arXiv, form interface, contact.