Natural Language Explanations (NLE) aim at explaining the decision-making process of a model through human-friendly natural text. Existing NLE approaches involve training separate models for each downstream task. In this work, we propose Uni-NLX, a unified framework that consolidates all NLE tasks into a single and compact multi-task model using a unified training objective of text generation. Additionally, we introduce two new NLE datasets: 1) ImageNetX, a dataset of 144K samples for explaining ImageNet categories, and 2) VQA-ParaX, a dataset of 123K samples for explaining the task of Visual Question Answering (VQA). Both datasets are derived leveraging large language models (LLMs). By training on the 1M combined NLE samples, our single unified framework is capable of simultaneously performing seven NLE tasks including VQA, visual recognition and visual reasoning tasks with 7$\times$ fewer parameters, demonstrating comparable performance to the independent task-specific models in previous approaches, and in certain tasks even outperforming them.
Sammani, F & Deligiannis, N 2023, Uni-NLX: Unifying Textual Explanations for Vision and Vision-Language Tasks. in VLAR, International Conference on Computer Vision Workshops (ICCVW) 2023. Workshop edn, vol. Workshop, IEEE, ICCV , pp. 1-4, VISION-AND-LANGUAGE ALGORITHMIC REASONING, ICCVW, Paris, France, 3/10/23.
Sammani, F., & Deligiannis, N. (2023). Uni-NLX: Unifying Textual Explanations for Vision and Vision-Language Tasks. In VLAR, International Conference on Computer Vision Workshops (ICCVW) 2023 (Workshop ed., Vol. Workshop, pp. 1-4). IEEE.
@inproceedings{f450c78ee3804dc0853d147dd7f53802,
title = "Uni-NLX: Unifying Textual Explanations for Vision and Vision-Language Tasks.",
abstract = "Natural Language Explanations (NLE) aim at explaining the decision-making process of a model through human-friendly natural text. Existing NLE approaches involve training separate models for each downstream task. In this work, we propose Uni-NLX, a unified framework that consolidates all NLE tasks into a single and compact multi-task model using a unified training objective of text generation. Additionally, we introduce two new NLE datasets: 1) ImageNetX, a dataset of 144K samples for explaining ImageNet categories, and 2) VQA-ParaX, a dataset of 123K samples for explaining the task of Visual Question Answering (VQA). Both datasets are derived leveraging large language models (LLMs). By training on the 1M combined NLE samples, our single unified framework is capable of simultaneously performing seven NLE tasks including VQA, visual recognition and visual reasoning tasks with 7$\times$ fewer parameters, demonstrating comparable performance to the independent task-specific models in previous approaches, and in certain tasks even outperforming them.",
author = "Fawaz Sammani and Nikos Deligiannis",
year = "2023",
language = "English",
volume = "Workshop",
pages = "1--4",
booktitle = "VLAR, International Conference on Computer Vision Workshops (ICCVW) 2023",
publisher = "IEEE",
edition = "Workshop",
note = "VISION-AND-LANGUAGE ALGORITHMIC REASONING, ICCVW, VLAR ; Conference date: 03-10-2023",
url = "https://wvlar.github.io/iccv23/",
}