Summary of Empirical Influence Functions to Understand the Logic Of Fine-tuning, by Jordan K. Matelsky et al.
Empirical influence functions to understand the logic of fine-tuning
by Jordan K. Matelsky, Lyle Ungar, Konrad P. Kording
First submitted to arxiv on: 1 Jun 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI)
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 A novel approach is proposed to understand the process of learning in neural networks, enabling improved performance and behavior interpretation. By analyzing the influence of fine-tuning on a new training sample, the paper identifies desiderata for such influences, including semantic distance, sparseness, noise invariance, transitive causality, and logical consistency. Empirical influence measured using fine-tuning is used to demonstrate how individual training samples affect outputs, revealing violations of these desiderata for both simple convolutional networks and modern Large Language Models (LLMs). The study also shows that prompting can partially rescue this failure. The paper presents a practical method for quantifying neural network learning from fine-tuning stimuli, suggesting popular models lack generalization capabilities and logical reasoning abilities. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Neural networks are super smart computers that learn from data, but we don’t fully understand how they do it. This research tries to figure out what happens when these networks learn from new information. The scientists want to know if the network’s output changes in a way that makes sense and is consistent with our understanding of language and logic. They found that most popular models can’t even perform simple logical tasks, which is surprising! The researchers also showed that by using special prompts, they could partially fix this problem. This study helps us understand how neural networks learn and what they’re capable of. |
Keywords
» Artificial intelligence » Fine tuning » Generalization » Neural network » Prompting
