SVD-based channel pruning for convolutional neural network in acoustic scene classification model

Convolutional Neural Network (CNN) offers promising performance in Acoustic Scene Classification (ASC) tasks. The CNN model, however, often involves a large number of parameters, and thus requires large storage space for the implementation of the model. In this paper, we propose a new method for model pruning based on singular value decomposition (SVD). More specifically, the number of parameters is reduced by a low-rank decomposition method, where a matrix is decomposed into products of three small matrices. As a result, the original convolutional layer is decomposed into three smaller convolutional layers resulting in an overall reduction in the number of parameters involved in the model. The proposed method is evaluated on the dataset of ASC task in DCASE2018. The results illustrate that the proposed approach dramatically reduces the CNN layers size by more than 90% with relatively 1% performance loss and the activity of parameters in convolutional layers increases with performance loss of the compressed model.