Summary of Structural Causality-based Generalizable Concept Discovery Models, by Sanchit Sinha et al.
Structural Causality-based Generalizable Concept Discovery Models
by Sanchit Sinha, Guangzhi Xiong, Aidong Zhang
First submitted to arxiv on: 20 Oct 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Methodology (stat.ME)
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 This paper proposes a novel disentanglement mechanism for learning mutually independent generative factors using Variational Autoencoders (VAEs) and structural causal models (SCMs). The method assumes a bipartite graph between generative factors and concepts, with directed edges from the former to the latter. Experiments are conducted on D-sprites and Shapes3D datasets, demonstrating successful learning of task-specific concepts explained by causal edges from generative factors. This approach generalizes well to arbitrary numbers of concepts and tasks. The paper’s contributions include a new method for disentangling concepts, which can be applied to various downstream tasks. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This research explores ways to explain how deep neural networks work. Right now, there are many different methods for understanding what these networks are doing. The main idea in this paper is to use two types of models: one that helps us understand the basic building blocks of data and another that shows us how these building blocks relate to each other. This combination allows us to figure out which concepts are most important for a specific task. The authors test their method on two different datasets and show that it works well. |
