Mitigating Overfitting for Deep Learning-based Aging-related Bug Prediction via Brain-inspired Regularization in Spiking Neural Networks

To alleviate the impact of software aging, primarily induced by aging-related bugs (ARBs), ARB prediction has drawn considerable interest from both academia and industry. Recent advances in deep learning (DL) have brought tremendous gains in ARB prediction. However, due to the limited size and extreme class imbalance in ARB datasets, conventional artificial neural networks (ANNs) are susceptible to overfitting, resulting in a suboptimal generalization performance. In this paper, we take advantage of sparse and binary nature of spiking communication in spiking neural networks (SNNs), which inherently provides a brain-inspired regularization to effectively alleviate overfitting. We propose the first spiking convolutional neural network-based ARB prediction model (ARB-SCNN), comprising a spiking encoder followed by a classifier and utilizing the Leaky Integrate-and-Fire neuron as the basic spiking computing unit. Considering the spatial-temporal dynamics and the non-differentiability nature, we develop a dedicated training framework for ARB-SCNN, which incorporates the rate coding-based mean square error (MSE) loss and employs the backpropagation through time with the surrogate gradient. Finally, extensive experiments on two real-world ARB datasets demonstrate that our ARB-SCNN effectively mitigates overfitting, improving generalization performance by 7.82% compared to the state-of-the-art DL-based classifiers, and it exhibits up to 5× better computational energy efficiency.