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 > Machine Learning

Title: A Koopman Approach to Understanding Sequence Neural Models

Abstract: We introduce a new approach to understanding trained sequence neural models: the Koopman Analysis of Neural Networks (KANN) method. Motivated by the relation between time-series models and self-maps, we compute approximate Koopman operators that encode well the latent dynamics. Unlike other existing methods whose applicability is limited, our framework is global, and it has only weak constraints over the inputs. Moreover, the Koopman operator is linear, and it is related to a rich mathematical theory. Thus, we can use tools and insights from linear analysis and Koopman Theory in our study. For instance, we show that the operator eigendecomposition is instrumental in exploring the dominant features of the network. Our results extend across tasks and architectures as we demonstrate for the copy problem, and ECG classification and sentiment analysis tasks.
Subjects: Machine Learning (cs.LG); Dynamical Systems (math.DS)
Cite as: arXiv:2102.07824 [cs.LG]
  (or arXiv:2102.07824v3 [cs.LG] for this version)

Submission history

From: Ilan Naiman [view email]
[v1] Mon, 15 Feb 2021 20:05:11 GMT (1208kb,D)
[v2] Wed, 10 Mar 2021 06:30:36 GMT (1209kb,D)
[v3] Thu, 7 Oct 2021 15:34:04 GMT (960kb,D)

Link back to: arXiv, form interface, contact.