An Explainability-driven Approach ForCompressing Convolutional Neural Networks

Convolutional neural networks (CNNs) havedemonstrated remarkable performance in various tasks, butthis often comes at the expense of high computational costs andmemory usage. Compression techniques, such as pruning andquantization, are applied to reduce the memory footprint ofCNNs and enable their deployment on resource-constrainededge devices. Recently, explainable artificial intelligence (XAI)methods have been introduced with the purpose ofunderstanding and explaining AI methods. XAI can be utilizedto gain insights into the inner workings of CNNs, such as thesignificance of different neurons and weights in the overallperformance of CNNs. This paper presents a unified approachfor compressing CNNs using explainable artificial intelligence(XAI). The proposed method calculates the importance scores ofCNN parameters through a low computing cost gradient-basedXAI technique called Layer-wise Relevance Propagation (LRP).These scores are then used to compress the CNN model; i.e.parameters with zero importance scores are pruned and theparameters with higher/lower scores are quantized withhigh/low number of bits. The proposed approach is evaluated ontwo CNN models (VGG۱۶ and AlexNet) on datasets such asFMNIST and CIFAR۱۰. The results show that the approachreduces the model size in convolutional layers by ۸۱% while onlydecreasing accuracy by around ۱۵% compared to the state-ofthe-art XAI-based compression method.

Keywords:

Convolutional Neural Network , Compression , Explainable-AI