Abstract
Deep convolutional neural networks have proven their effectiveness and have been acknowledged as the most dominant method for image classification. However, a severe drawback of deep convolutional neural networks is poor explainability. Unfortunately, in many real-world applications, users need to understand the rationale behind the predictions of deep convolutional neural networks when determining whether they should trust the predictions or not. To resolve this issue, a novel genetic algorithm-based method is proposed for the first time to automatically evolve local explanations that can assist users to assess the rationality of the predictions. Furthermore, the proposed method is model-agnostic, i.e., it can be utilised to explain any deep convolutional neural network models. 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 characteristic of the proposed method. The evolved local explanations on four images, randomly selected from ImageNet, are presented, which show that the evolved local explanations are straightforward to be recognised by humans. Moreover, the evolved explanations can explain the predictions of deep convolutional neural networks on all four images very well by successfully capturing meaningful interpretable features of the sample images. Further analysis based on the 30 runs of the experiments exhibits that the evolved local explanations can also improve the probabilities/confidences of the deep convolutional neural network models in making the predictions. The proposed method can obtain local explanations within one minute, which is more than ten times faster than LIME (the state-of-the-art method).
Introduction
Deep Convolutional Neural Networks (CNNs) are often criticized as "black boxes" due to their complex internal structures. While they achieve state-of-the-art accuracy in image classification, the lack of transparency is a major hurdle for their adoption in high-stakes environments like medicine or autonomous driving. Understanding why a model classifies an image as a specific class is essential for building user trust.
Existing methods like LIME (Local Interpretable Model-agnostic Explanations) provide local explanations by approximating a complex model with a simpler, interpretable one. However, LIME suffers from significant computational overhead due to its random sampling process and requires manual tuning of the number of features to include in an explanation. This research introduces E-LIME, which leverages Genetic Algorithms (GA) to overcome these limitations.
The E-LIME Methodology
E-LIME treats the generation of an explanation as a combinatorial optimization problem. The process begins with image segmentation using the SLIC algorithm to create "superpixels," which serve as the basic units of interpretation.
As shown in the evolutionary process, candidate explanations are encoded as binary vectors representing the presence or absence of specific superpixels. A Genetic Algorithm then evolves these candidates. The fitness function is the core of this approach: it measures the probability or confidence assigned by the original "black-box" CNN to the target class when given the masked image. By maximizing this confidence, the GA naturally identifies the most critical features while discarding irrelevant background noise.
Experimental Evaluation
The researchers evaluated E-LIME using the ImageNet dataset across several popular architectures to demonstrate its model-agnostic nature.
Models tested included ResNet-34, DenseNet-121, and MobileNetV2. The experiments focused on three key areas: the visual clarity of the explanations, the impact on model confidence, and the computational speed compared to traditional LIME.
Key Results and Discussion
The experimental results highlight several advantages of the E-LIME approach:
- Significant Speedup: E-LIME generates explanations in under one minute, making it more than ten times faster than standard LIME. This efficiency stems from the GA's ability to focus the search on promising regions of the solution space rather than relying on thousands of random perturbations.
- Confidence Optimization: Interestingly, the "evolved" explanations (masked images) often resulted in higher classification probabilities than the original images. For example, on a "Samoyed" image, the probability increased from 0.9574 to 0.9988 by removing distracting background elements.
- Intuitive Explanations: The method successfully isolated critical features, such as the "red eye" of a tench fish or the "hooked beak" of a bald eagle. These results align well with human intuition, making the model's decision-making process transparent.
Conclusion
E-LIME represents a significant step forward in making deep learning models more interpretable. By replacing the expensive sampling and manual tuning of LIME with an evolutionary search, the proposed method offers a fast, automated, and model-agnostic solution for generating local explanations. Future work will explore more rigorous quantitative metrics for faithfulness and extend the GA-based approach to other data modalities such as text and audio.
Summary updated