The impact of pruning Convolutional Neural Networks when classifying skin cancer

Abstract: Over the past few years, there have been multiple reports showcasing how Convolutional Neural Networks (CNNs) can be used to classify if skin lesions are cancerous or non-cancerous. However, a limitation of CNNs is the large number of parameters resulting in high computation times. One method to reduce the computation times is using pruning to remove some parameters. This report therefore aims to answer: ”How does pruning impact the accuracy and inference time when identifying skin cancer pictures from the HAM10000 dataset using the VGG-16 architecture?” The CNN model VGG-16 was trained and tested for ten different levels of pruning, 0-90% with intervals of 10%. Due to how pruning is implemented in TensorFlow there is no measurable speedup in inference time so it was estimated by removing entire convolutional layers. The results show no significant change in accuracy between 0-60% pruning, a slight drop in accuracy at 70% pruning, and a large accuracy drop when pruning 80-90% of the parameters. The results also showed that the inference time could be made almost 50% faster by removing 60% of the parameters. Thus it is concluded that pruning can be a viable option to reduce interference time in a CNN.