Human-in-the-Loop: Active Learning Approach to Image Classification using CLIP

Keywords

No Thumbnail Available

Issue Date

2023-07-11

Language

en

Document type

Journal Title

Journal ISSN

Volume Title

Publisher

Title

ISSN

Volume

Issue

Startpage

Endpage

DOI

Abstract

Image classification can be a labor-intensive job, because of manually labeling the data. Previous research within Squadra Machine Learning Company has focused on the capabilities of large multimodal models for image classification and feature extraction. It was found that the zero-shot capabilities of these models do not always suffice and fine-tuning is necessary. This paper investigates the possibilities of fine-tuning an image model for few-shot classification. For this, a human-in-the-loop Active Learning approach was used to train OpenAI’s CLIP model without a highly labor-intensive labeling task. The research question that will be tried to answer is: “What is the impact of different uncertainty measures on the performance of Active Learning in few-shot classification tasks, and how does this approach compare to traditional machine learning?” The report discusses the four main uncertainty measures and compares the performance of Active Learning to traditional machine learning. Results show that entropy sampling and margin outperform traditional machine learning and are potential approaches to use at Squadra Machine Learning Company

Description

Citation

Supervisor

Faculty

Faculteit der Sociale Wetenschappen