In many types of music, percussion plays an essential role to establish the rhythm and the groove of the music. Algorithms that can decompose the percussive signal into its constituent components would therefore be very useful, as they would enable many analytical and creative applications. This paper describes a method for the unsupervised decomposition of percussive recordings, building on the non-negative matrix factor deconvolution (NMFD) algorithm. Given a percussive music recording, NMFD discovers a dictionary of time-varying spectral templates and corresponding activation functions, representing its constituent sounds and their positions in the mix. We observe, however, that the activation functions discovered using NMFD do not show the expected impulse-like behavior for percussive instruments. We therefore enforce this behavior by specifying that the activations should take on binary values: either an instrument is hit, or it is not. To this end, we rewrite the activations as the output of a sigmoidal function, multiplied with a per-component amplitude factor. We furthermore define a regularization term that biases the decomposition to solutions with saturated activations, leading to the desired binary behavior. We evaluate several optimization strategies and techniques that are designed to avoid poor local minima. We show that incentivizing the activations to be binary indeed leads to the desired impulse-like behavior, and that the resulting components are better separated, leading to more interpretable decompositions.
Deep neural networks have achieved state-of-the-art performance in image classification. Due to this success, deep learning is now also being applied to other data modalities such as multispectral images, lidar and radar data. However, successfully training a deep neural network requires a large reddataset. Therefore, transitioning to a new sensor modality (e.g., from regular camera images to multispectral camera images) might result in a drop in performance, due to the limited availability of data in the new modality. This might hinder the adoption rate and time to market for new sensor technologies. In this paper, we present an approach to leverage the knowledge of a teacher network, that was trained using the original data modality, to improve the performance of a student network on a new data modality: a technique known in literature as knowledge distillation. By applying knowledge distillation to the problem of sensor transition, we can greatly speed up this process. We validate this approach using a multimodal version of the MNIST dataset. Especially when little data is available in the new modality (i.e., 10 images), training with additional teacher supervision results in increased performance, with the student network scoring a test set accuracy of 0.77, compared to an accuracy of 0.37 for the baseline. We also explore two extensions to the default method of knowledge distillation, which we evaluate on a multimodal version of the CIFAR-10 dataset: an annealing scheme for the hyperparameter and selective knowledge distillation. Of these two, the first yields the best results. Choosing the optimal annealing scheme results in an increase in test set accuracy of 6%. Finally, we apply our method to the real-world use case of skin lesion classification.
10 January 2018 in Neural Computing and Applications