Explaining Deep Convolutional Neural Networks for Image Classification by Evolving Local Interpretable Model-Agnostic Explanations

Bin Wang, Wenbin Pei, Bing Xue, Mengjie Zhang

Abstract: Deep convolutional neural networks (CNNs) have proven their effectiveness and are widely acknowledged as the dominant method for image classification. However, their lack of explainability remains a significant drawback, particularly in real-world applications where users need to understand the rationale behind predictions to determine their trustworthiness. Local Interpretable Model-agnostic Explanations (LIME) is a popular method for explaining deep CNN predictions, but it suffers from two major limitations: (1) a computationally expensive sampling process to generate perturbed images, and (2) the need to pre-define the number of interpretable features (superpixels), which often requires manual fine-tuning. To address these limitations, we propose a novel Genetic Algorithm (GA)-based method, called E-LIME, which automatically evolves local explanations for deep CNNs. The proposed method eliminates the need for the computationally expensive sampling process used in LIME and allows for the automatic selection of superpixels without pre-defining their number. Specifically, E-LIME introduces a flexible encoding strategy to represent superpixels as binary vectors and a new fitness function that evaluates the selected superpixels based on the probability of the deep CNN making a specific prediction. By optimising the fitness value, E-LIME selects the most informative superpixels while removing noisy features, resulting in more efficient and accurate local explanations. In the experiments, ResNet is used as an example model to be explained, and the ImageNet dataset is selected as the benchmark dataset. DenseNet and MobileNet are further explained to demonstrate the model-agnostic characteristics of the proposed method. The evolved local explanations on four randomly selected images from ImageNet show that the proposed method successfully captures meaningful interpretable features, improving the probabilities/confidences of the deep CNN models in making predictions. Moreover, the proposed method obtains local explanations within one minute, which is more than ten times faster than LIME. The proposed E-LIME method not only overcomes the limitations of LIME but also provides a more efficient and flexible approach to explaining deep CNN predictions, making it highly suitable for real-world applications where interpretability and computational efficiency are critical.

External IDs:dblp:journals/tetci/WangPXZ25