An Improved Analysis Of TrainingOver-paramet Erized Deep Neural Networks.pdf

8479-an-improved-analysis-of-training-over-parameterized-deep-neural-networks.pdf
Preview of An Improved Analysis of TrainingOver-paramet erized Deep Neural Networks
🔗 Source: papers.nips.cc
📊 Size: 491 KB
👤 Author: Difan Zou, Quanquan Gu
⬇️ Downloads: 83

Summary

The authors provide a milder over-parameterization condition and faster global convergence rates than previous work. They achieve this through two innovative proof techniques: a tighter gradient lower bound and a sharper characterization of the trajectory length. The results show that with Gaussian random initialization, gradient descent can achieve ε training loss within a certain number of iterations, and stochastic gradient descent can achieve ε expected training loss within a certain number of iterations. The over-parameterization condition is milder by a factor of eΩ(n16φ−4) and the iteration complexity is better by a factor of eO(n4φ−1) compared to the state-of-the-art result. The authors also specialize their results to two-layer ReLU networks, which outperforms the best-known results.

Description

The authors provide a milder over-parameterization condition and faster global convergence rates than previous work.

Technical Information

  • File Format: PDF
  • File Size: 491 KB
  • Pages: 10
  • Language: EN
  • Author: Difan Zou, Quanquan Gu
  • Total Downloads: 83
  • Last Updated: 13 hours ago

Document Overview

This PDF document about An Improved Analysis of TrainingOver-paramet erized Deep Neural Networks provides comprehensive information and guidance. Whether you're a beginner or advanced user, this resource offers valuable insights into An Improved Analysis of TrainingOver-paramet erized Deep Neural Networks.

Related Topics

If you're interested in An Improved Analysis of TrainingOver-paramet erized Deep Neural Networks, you might also want to explore:

Download An Improved Analysis of TrainingOver-paramet erized Deep Neural Networks eBooks for free and learn more about An Improved Analysis of TrainingOver-paramet erized Deep Neural Networks. These books contain exercises and tutorials to improve your practical skills, at all levels!

Not satisfied with this document? We have related documents to An Improved Analysis of TrainingOver-paramet erized Deep Neural Networks, try searching with similar keywords: An Improved Analysis of TrainingOver-paramet erized Deep Neural Networks, Neural Networks R Rojas Neural Networks Springer V, Models Of Neural Networks Physics Of Neural Networ, Neural Networks Evolutionary Swarm Neural Network , Improved Quality Improved Profits, Cara Menghitung Jumlah Dengan Lebih Dari 3 Paramet, Practical Diagnostic Tools for Deep Neural Networks, Free Pdf Epub Download Introduction To Neural Networks Using Matlab 6 0 S N Sivanandam S N Deep

You can download PDF versions of the user's guide, manuals and ebooks about An Improved Analysis of TrainingOver-paramet erized Deep Neural Networks, you can also find and download for free A free online manual (notices) with beginner and intermediate, Downloads Documentation, You can download PDF files (or DOC and PPT) about An Improved Analysis of TrainingOver-paramet erized Deep Neural Networks for free, but please respect copyrighted ebooks.