Summary of Curriculum Learning with Quality-driven Data Selection, by Biao Wu et al.
Curriculum Learning with Quality-Driven Data Selection
by Biao Wu, Fang Meng, Ling Chen
First submitted to arxiv on: 27 Jun 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
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 The proposed methodology in this paper aims to improve the development of Multimodal Large Language Models (MLLMs) by enhancing zero-shot capabilities across various tasks. The authors utilize visual instruction tuning with machine-generated instruction-following data, which has shown to boost performance in different domains. However, existing methodologies for selecting MLLM training data often rely on single, unreliable scores or use downstream tasks, leading to potential overfitting and time-consuming evaluation processes. To address these limitations, the researchers introduce a novel approach that leverages image-text correlation and model perplexity to evaluate and select data of varying quality. This method maps data quality into a two-dimensional space, enabling the selection of high-quality data based on their location within this distribution. The authors demonstrate the effectiveness of their methodology by analyzing the impact of task type settings (prompts) on data quality and constructing multi-stage subsets for curriculum learning. Comprehensive experiments are conducted on various datasets, showcasing substantial enhancements in five commonly assessed capabilities compared to using the complete dataset. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper proposes a new way to train Multimodal Large Language Models. Right now, people use machine-generated instructions to make these models better at doing tasks without being taught. The problem is that this method can be unreliable and takes a long time. To solve this issue, the researchers created a new system that looks at how well images and text match up and how well a model understands what it’s reading. This helps choose the best data to train the models with. They tested their method on different datasets and showed that it makes the models much better at doing tasks without being taught. |
Keywords
» Artificial intelligence » Curriculum learning » Instruction tuning » Overfitting » Perplexity » Zero shot
