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

Download:

Current browse context:

cs.LG

Change to browse by:

cs

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: Magnitude and Angle Dynamics in Training Single ReLU Neurons

Abstract: To understand learning the dynamics of deep ReLU networks, we investigate the dynamic system of gradient flow $w(t)$ by decomposing it to magnitude $w(t)$ and angle $\phi(t):= \pi - \theta(t) $ components. In particular, for multi-layer single ReLU neurons with spherically symmetric data distribution and the square loss function, we provide upper and lower bounds for magnitude and angle components to describe the dynamics of gradient flow. Using the obtained bounds, we conclude that small scale initialization induces slow convergence speed for deep single ReLU neurons. Finally, by exploiting the relation of gradient flow and gradient descent, we extend our results to the gradient descent approach. All theoretical results are verified by experiments.
Subjects: Machine Learning (cs.LG)
Cite as: arXiv:2209.13394 [cs.LG]
  (or arXiv:2209.13394v2 [cs.LG] for this version)

Submission history

From: Jong Chul Ye [view email]
[v1] Tue, 27 Sep 2022 13:58:46 GMT (1097kb,D)
[v2] Wed, 12 Oct 2022 02:14:20 GMT (1097kb,D)

Link back to: arXiv, form interface, contact.