Deep Model Interpretation with Limited Data : A Coreset-based Approach

TL;DR

A coreset-based interpretation framework is proposed that utilizes coreset selection methods to sample a representative subset of the large dataset for the interpretation task and a similarity-based evaluation protocol is proposed to assess the robustness of model interpretation methods towards the amount data they take as input.

Abstract

Model Interpretation aims at the extraction of insights from the internals of a trained model. A common approach to address this task is the characterization of relevant features internally encoded in the model that are critical for its proper operation. Despite recent progress of these methods, they come with the weakness of being computationally expensive due to the dense evaluation of datasets that they require. As a consequence, research on the design of these methods have focused on smaller data subsets which may led to reduced insights. To address these computational costs, we propose a coreset-based interpretation framework that utilizes coreset selection methods to sample a representative subset of the large dataset for the interpretation task. Towards this goal, we propose a similarity-based evaluation protocol to assess the robustness of model interpretation methods towards the amount data they take as input. Experiments considering several interpretation methods, DNN models, and coreset selection methods show the effectiveness of the proposed framework.

Figures (8)

• Figure 1: Proposed is a coreset-based model interpretation framework. The CNN model computes activation maps for the entire dataset, then a coreset selection method picks representative subsets. Using these subsets and entire activation maps, an interpretation method obtains relevant internal features encoded by the CNN model. Finally, two evaluations compare the quality of obtained features from the coreset.
• Figure 2: Classification accuracy based on relevant features by different combinations of interpretation and coreset selection methods.
• Figure 3: An analysis of identified units by VEBI using different coreset selection methods on Resnet18. Left shows number of identified units per layer. Right shows coverage of similar identified units between each coreset selection method and complete dataset. Rand, DGP, and MS stand for Random, DGPruning, and Moderate Selection methods.
• Figure 4: Relevant features similarity between corset and complete dataset for interpretation methods ICE (left), Topic-based interpretation (middle), VEBI (right) over VGG19, Resnet18, and Resnet50.
• Figure 5: Relevant feature similarity using coreset transferability, i.e., corsets determined by Resnet50, for interpretation methods ICE (top) and VEBI (down) over VGG19 and Resnet18.