Identification and Uses of Deep Learning Backbones via Pattern Mining
This addresses the interpretability gap in deep learning for data mining applications, offering a method to analyze and enhance model behavior, though it is incremental as it builds on existing pattern mining and ILP techniques.
The paper tackles the problem of understanding deep learning predictions by identifying a 'backbone' of neurons associated with specific groups of instances, such as classes or misclassifications, and demonstrates its use for improving performance, explanation, and visualization on datasets like Bird Audio Detection, Labeled Faces in the Wild, and MNIST.
Deep learning is extensively used in many areas of data mining as a black-box method with impressive results. However, understanding the core mechanism of how deep learning makes predictions is a relatively understudied problem. Here we explore the notion of identifying a backbone of deep learning for a given group of instances. A group here can be instances of the same class or even misclassified instances of the same class. We view each instance for a given group as activating a subset of neurons and attempt to find a subgraph of neurons associated with a given concept/group. We formulate this problem as a set cover style problem and show it is intractable and presents a highly constrained integer linear programming (ILP) formulation. As an alternative, we explore a coverage-based heuristic approach related to pattern mining, and show it converges to a Pareto equilibrium point of the ILP formulation. Experimentally we explore these backbones to identify mistakes and improve performance, explanation, and visualization. We demonstrate application-based results using several challenging data sets, including Bird Audio Detection (BAD) Challenge and Labeled Faces in the Wild (LFW), as well as the classic MNIST data.