Summary of Why Are Visually-grounded Language Models Bad at Image Classification?, by Yuhui Zhang et al.
Why are Visually-Grounded Language Models Bad at Image Classification?
by Yuhui Zhang, Alyssa Unell, Xiaohan Wang, Dhruba Ghosh, Yuchang Su, Ludwig Schmidt, Serena Yeung-Levy
First submitted to arxiv on: 28 May 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary Revisiting image classification with visually-grounded language models (VLMs) like GPT-4V and LLaVA reveals that existing proprietary and public VLMs underperform CLIP on standard benchmarks like ImageNet, despite using similar architectures. Our analysis points to data-related issues as the primary cause: VLMs’ latent spaces encode critical information for image classification, but can only be effectively decoded with sufficient training data. A strong correlation exists between class exposure during training and instruction-tuning and a VLM’s performance in those classes. When trained with enough data, VLMs match state-of-the-art classification model accuracy. To overcome this limitation, we integrate classification-focused datasets into the VLM’s training, enhancing its performance by 11.8% on the newly collected ImageWikiQA dataset. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about how language models can be used for image classification tasks. We found that some language models don’t perform well on this task even though they have many parameters and use similar tools as other good models. We looked into why this might be happening and discovered that it’s because these language models haven’t been trained on enough images to learn how to recognize objects. When we train one of these language models with more data, it becomes much better at classifying images. This is important because it means we can use language models for even more tasks in the future. |
Keywords
» Artificial intelligence » Classification » Gpt » Image classification » Instruction tuning
