Smart-Init of neural networks

David Denisov, Dan Feldman, Shlomi Dolev, Michael Segal

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

The initialization of neural networks is of significant importance for their performance. Currently, the prevalent initialization method is a random sample based on the network’s structure. This work presents a general yet effective method to initialize neural networks. We provide repeated experiments of training variants of Mobile-Net over down-sampled variants of Image-Net, demonstrating accuracy gain and loss decrease across most of the test sets and validation sets. E.g., for Mobile-Net (v1) and Image-Net (32 × 32), we had 2.5% accuracy improvement over both the test and validation sets.

Original languageEnglish
Title of host publicationProceedings - 2025 11th International Conference on Computing and Artificial Intelligence, ICCAI 2025
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages250-255
Number of pages6
ISBN (Electronic)9798331524913
DOIs
StatePublished - 2025
Event11th International Conference on Computing and Artificial Intelligence, ICCAI 2025 - Kyoto, Japan
Duration: 28 Mar 202531 Mar 2025

Publication series

NameProceedings - 2025 11th International Conference on Computing and Artificial Intelligence, ICCAI 2025

Conference

Conference11th International Conference on Computing and Artificial Intelligence, ICCAI 2025
Country/TerritoryJapan
CityKyoto
Period28/03/2531/03/25

Bibliographical note

Publisher Copyright:
©2025 IEEE.

Keywords

  • Neural networks
  • Neural networks initialization
  • Non-convex optimization
  • Optimization

ASJC Scopus subject areas

  • Computer Science Applications
  • Control and Systems Engineering
  • Computer Graphics and Computer-Aided Design
  • Artificial Intelligence

Fingerprint

Dive into the research topics of 'Smart-Init of neural networks'. Together they form a unique fingerprint.

Cite this