TransBoost: Improving the Best ImageNet Performance using Deep Transduction

TL;DR

TransBoost is a simple, efficient transductive fine-tuning method that significantly enhances deep neural network performance on ImageNet and other image classification datasets, achieving state-of-the-art results.

Contribution

Proposes TransBoost, a novel transductive fine-tuning procedure that improves deep neural network performance on unlabeled test sets across various architectures.

Findings

Significant performance improvements on ImageNet classification.

Effective across multiple neural network architectures.

Achieves state-of-the-art transductive results.

Abstract

This paper deals with deep transductive learning, and proposes TransBoost as a procedure for fine-tuning any deep neural model to improve its performance on any (unlabeled) test set provided at training time. TransBoost is inspired by a large margin principle and is efficient and simple to use. Our method significantly improves the ImageNet classification performance on a wide range of architectures, such as ResNets, MobileNetV3-L, EfficientNetB0, ViT-S, and ConvNext-T, leading to state-of-the-art transductive performance. Additionally we show that TransBoost is effective on a wide variety of image classification datasets. The implementation of TransBoost is provided at: https://github.com/omerb01/TransBoost .