核心概念
Understanding and utilizing deep learning backbones for improved performance and explanation.
摘要
The content explores the identification and application of deep learning backbones through pattern mining. It delves into the core idea of concept backbones and their significance in improving model performance and understanding. The paper presents a heuristic approach to finding these backbones efficiently and demonstrates their application across various datasets. The experiments showcase the effectiveness of the method in identifying mistakes, improving predictions, and providing visual explanations.
統計資料
Deep learning is used as a black-box method with impressive results.
Identifying a backbone of deep learning for a given group of instances is explored.
The problem is formulated as a set cover style problem and shown to be intractable.
A coverage-based heuristic approach related to pattern mining is explored.
Backbones are used to identify mistakes, improve performance, explanation, and visualization.
引述
"A core insight is that any instance activates a subset of neurons in the network."
"Our approach is flexible enough to answer a variety of questions."