Classification of musical instruments with convolutional neural networks

This paper presents one solution to a problem of classifying musical instruments with convolutional neural networks. Mel frequency cepstral coefficients are used for audio features extraction, and neural network architecture is modelled after the LeNet architecture. During the learning process, Adam optimization is used, along with negative log likelihood loss function. In the end, results are given and it is concluted that this solution has, at least 4% better accuracy on the validation set, than any other published solution.