Declaration-based Prompt Tuning for Visual Question Answering

TL;DR

This paper introduces Declaration-based Prompt Tuning (DPT), a novel approach that unifies pre-training and fine-tuning objectives for VQA, significantly improving accuracy and generalization, especially in low-data scenarios.

Contribution

DPT reformulates VQA as a prompt-tuning task with joint optimization of pre-training and fine-tuning objectives, enhancing model adaptation and performance.

Findings

DPT outperforms traditional fine-tuning by 2.68% in accuracy on GQA.

DPT achieves over 31% improvement in zero-shot and few-shot settings.

Experimental results demonstrate DPT's effectiveness in both fully-supervised and low-data scenarios.

Abstract

In recent years, the pre-training-then-fine-tuning paradigm has yielded immense success on a wide spectrum of cross-modal tasks, such as visual question answering (VQA), in which a visual-language (VL) model is first optimized via self-supervised task objectives, e.g., masked language modeling (MLM) and image-text matching (ITM), and then fine-tuned to adapt to downstream task (e.g., VQA) via a brand-new objective function, e.g., answer prediction. The inconsistency of the objective forms not only severely limits the generalization of pre-trained VL models to downstream tasks, but also requires a large amount of labeled data for fine-tuning. To alleviate the problem, we propose an innovative VL fine-tuning paradigm (named Declaration-based Prompt Tuning, abbreviated as DPT), which jointly optimizes the objectives of pre-training and fine-tuning of VQA model, boosting the effective adaptation of pre-trained VL models to the downstream task. Specifically, DPT reformulates the objective form of VQA task via (1) textual adaptation, which converts the given questions into declarative sentence-form for prompt-tuning, and (2) task adaptation, which optimizes the objective function of VQA problem in the manner of pre-training phase. Experimental results on GQA dataset show that DPT outperforms the fine-tuned counterpart by a large margin regarding accuracy in both fully-supervised (2.68%) and zero-shot/few-shot (over 31%) settings. All the data and codes will be available to facilitate future research.