Customizable and Programmable Deep Learning

Abstract

In this study, we explore the potential of pre-trained deep learning models, proposing a new approach that emphasizes their reusability and adaptability. Our framework, termed “customizable” deep learning, facilities users to seamlessly integrate diverse pre-trained models for addressing new tasks and enhancing existing solutions. Furthermore, we introduce a “programmable” adapter that enables the flexible combination of different pre-trained modules, expanding the range of applications and customization options. Through empirical experiments, particularly focusing on Visual Question Answering (VQA) for visually impaired (VI) individuals, we demonstrate the practical effectiveness of our methodology. These contributions advance the deep learning field while promoting customization and re-usability across various domains and tasks. The code is available https://github.com/Ratnabali-Pal/CPDA-VQA.