Explainable image recognition with graph-based feature extraction

Abstract

Deep learning models have proven remarkably adept at extracting salient features from raw data, driving state-of-the-art performance across many domains. However, these models suffer from a lack of interpretability; they function as black boxes, obscuring the feature-level support of their predictions. Addressing this problem, we introduce a novel framework that combines the strengths of convolutional layers in extracting features with the adaptability of Graph Neural Networks (GNNs) to effectively represent the interconnections among neuron activations. Our framework operates in two phases: first, it identifies class-oriented neuron activations by analyzing image features, then these activations are encapsulated within a graph structure. The GNN in our system utilizes the connections between neuron activations to yield an interpretable final classification. This approach allows for the backtracking of predictions to identify key contributing neurons, enhancing the model's explainability. The proposed model not only matches, but at times exceeds, the accuracy of current leading models, all the while providing transparency via class-specific feature importance. This novel integration of convolutional and graph neural networks offers a significant step towards interpretable and accountable deep learning models.