Software engineering efforts increasingly focus on enhancing software quality by detecting, addressing, and predicting defect-prone components. This area has garnered considerable research interest due to its integral role in the software development process. Numerous methods have been proposed for software defect prediction, with recent work emphasizing data mining strategies, especially the use of neural networks, as powerful tools. Despite their effectiveness, neural networks are often criticized for their opacity, leading to doubts about the reliability of their outcomes. To address this, we use DeepDebugger, an explainable AI technique designed for defect prediction. It leverages animated visualizations composed of dynamic, color-coded elements to represent the model’s learning behavior. We evaluated DeepDebugger across widely used neural network models on datasets from NASA, PROMISE, and AEEEM. Experimental findings show that this approach improves both the interpretability and the visualization of neural network training dynamics.