Complexity Reduction for Classification of Musical Instruments Using Element Selection

In this study, we propose complexity reduction for convolutional neural network (CNN)-based music instrument classification by incorporating an element selection method. This method reduces the dimension of input features to decrease computational costs, which is crucial for applications such as music genre identification and musical information retrieval, particularly in real-time systems. Unlike other dimensionality reduction techniques such as principal component analysis, our approach does not require multiplication, further lowering computational costs. The selected elements are optimized to minimize the mean reconstruction error, thus preserving estimation performance. Experimental evaluations demonstrated that our method outperformed other methods such as increasing the stride width or dilation rate in CNNs.