Summary of Training Objective Drives the Consistency Of Representational Similarity Across Datasets, by Laure Ciernik and Lorenz Linhardt and Marco Morik and Jonas Dippel and Simon Kornblith and Lukas Muttenthaler
Training objective drives the consistency of representational similarity across datasets
by Laure Ciernik, Lorenz Linhardt, Marco Morik, Jonas Dippel, Simon Kornblith, Lukas Muttenthaler
First submitted to arxiv on: 8 Nov 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI); 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 The Platonic Representation Hypothesis proposes that foundation models converge to a shared representation space based on downstream task performance, regardless of training objectives or data modalities. While representational similarity is typically measured for individual datasets, it’s unclear whether this convergence is dataset-specific. This paper explores how model representations vary depending on the stimuli used to construct them, finding that objective functions have a significant impact on consistency across datasets. Self-supervised vision models exhibit better generalization of pairwise similarities from one dataset to another compared to image classification or image-text models. The correspondence between representational similarity and task behavior is also dataset-dependent, being most pronounced for single-domain datasets. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Recent research suggests that foundation models are converging to a shared representation space based on their performance in different tasks. But how does this convergence work? Is it specific to certain types of data or tasks? This study looks at how model representations change depending on the type of data used to train them, and finds that the goal of training has a big impact on whether models learn similar or different things. They also find that some types of models are better than others at learning generalizable representations. |
Keywords
» Artificial intelligence » Generalization » Image classification » Self supervised
