Billion-scale semi-supervised learning for image classification (1905.00546v1)

Abstract: This paper presents a study of semi-supervised learning with large convolutional networks. We propose a pipeline, based on a teacher/student paradigm, that leverages a large collection of unlabelled images (up to 1 billion). Our main goal is to improve the performance for a given target architecture, like ResNet-50 or ResNext. We provide an extensive analysis of the success factors of our approach, which leads us to formulate some recommendations to produce high-accuracy models for image classification with semi-supervised learning. As a result, our approach brings important gains to standard architectures for image, video and fine-grained classification. For instance, by leveraging one billion unlabelled images, our learned vanilla ResNet-50 achieves 81.2% top-1 accuracy on the ImageNet benchmark.

• The paper presents a scalable semi-supervised framework that effectively uses billions of unlabeled images to improve classification.
• It demonstrates that integrating large-scale unlabeled data significantly boosts model performance with minimal additional labeling.
• The methodology sets new state-of-the-art benchmarks, highlighting practical implications for real-world image recognition tasks.

Overview of ICCV \LaTeX\ Author Guidelines

This document serves as a comprehensive guide for authors preparing manuscripts for the International Conference on Computer Vision (ICCV) proceedings using \LaTeX. It outlines essential formatting and submission requirements, ensuring that submissions adhere to ICCV's standards for consistency and professionalism.

Key Instructions

• Language and Submission Policies: Manuscripts must be composed in English. Authors are encouraged to consult the ICCV website for official policies regarding dual submissions and blind review processes.
• Length and Formatting: Papers are limited to eight pages, excluding references. The guide strictly prohibits overlength submissions, adhering to standard formatting for text and figures to ensure uniformity across submissions.
• Blind Review Process: The guide emphasizes the importance of maintaining anonymity. Authors should refrain from explicitly identifying themselves within the text or citations, while ensuring previous work by the authors is appropriately referenced.

Detailed Formatting Specifications

• Text Layout: Submissions must follow a two-column format with precise dimensions. Fonts and type styles are specified for various sections, including titles, author names, abstracts, and main text, to maintain visual coherence.
• Figures and Tables: Authors are instructed on how to incorporate graphics effectively. Captions should be clear and concise, with figures resized to harmonize with the document's text. The use of color is also addressed, with further details available on the ICCV website.
• Mathematical Notation and References: All mathematical equations should be numbered for ease of reference. A standardized citation style is mandated, with bibliographic references listed numerically at the document's conclusion.

Compliance and Submission

Authors are required to submit a signed IEEE copyright release form for their papers to be published. This document underscores the necessity for strict adherence to formatting rules to facilitate a streamlined review process and uphold the academic integrity of the proceedings.

Implications and Future Directions

While this document serves as a technical manual, its implications extend to the broader context of academic publishing in computer science. By enforcing standardized submission guidelines, ICCV maintains the quality and accessibility of its contributions, setting a precedent for careful attention to detail in academic writing. This rigorous approach ensures that ideas are communicated effectively and equitably assessed.

Ongoing developments in typesetting and document preparation could further enhance the submission process. Automation tools and advanced document editors may play an increasingly significant role in simplifying compliance with publication standards, potentially reducing the burden on authors and manuscript reviewers alike.