Given the huge size of music collections available on the Web, automatic genre classification is crucial for the organization, search, retrieval and recommendation of music. Different kinds of features have been employed as input to classification models which have been shown to achieve high accuracy in classification scenarios under controlled environments. In this work, we investigate two components of the music genre classification process: a novel feature vector obtained directly from a description of the musical structure described in MIDI files (named as structural features), and the performance of relational classifiers compared to the traditional ones. Neither structural features nor relational classifiers have been previously applied to the music genre classification problem. Our hypotheses are: (i) the structural features provide a more effective description than those currently employed in automatic music genre classification tasks, and (ii) relational classifiers can outperform traditional algorithms, as they operate on graph models of the data that embed information on the similarity between music tracks. Results from experiments carried out on a music dataset with unbalanced distribution of genres indicate these hypotheses are promising and deserve further investigation.

We propose a novel approach based on Deep Learning methods using a convolutional neural networks (CNNs) to compare observed four-dimensional (4-D) seismic data with reservoir simulation model results and select the simulation models with the best match. Our approach advantages are twofold: 1) it is not a pixel-based approach, as it learns spatial features through training examples and 2) it does not require the simulation model and 4-D seismic maps to be in the same domain, enabling a direct comparison of maps in the seismic domain with maps in the numerical simulation domain. Additionally, we present a new methodology to quantitatively assess the performance of each method relative to the expected results. For this purpose, we build an annotated dataset containing the maps that best match the 4-D seismic in different realistic scenarios while taking into account the specialists’ knowledge. We apply our methodology to evaluate if a method is able to perform according to what specialists expect, considering the annotated data. We also propose an improvement to traditional methods using a statistical method to convert maps of different physical properties into a common property, which can be used as an alternative to avoid forward and/or inversion procedures commonly used to bring 4-D seismic and simulation data to the same domain. Finally, we present an extensive comparison, qualitative and quantitative, of our approaches with other methods in the literature, such as the traditional pixelwise strategy for maps in the same domain, and methods based on binarization of maps in different domains. The results show that our proposed methods are capable of more accurately identifying the best-matched models, according to the specialists’ answers.

Users interact a lot with their personal music collections, typically using standard text-based interfaces that offer constrained functionalities based on assigned metadata or tags. Alternative visual interfaces have been developed, both to display graphical views of music collections that attempt to reflect some chosen property or organization, or to display abstract visual representations of specific songs. Yet, there are many dimensions involved in the perception and handling of music and mapping musical information into computer tractable models is a challenging problem. With a wide variety of possible approaches, the search for novel strategies to visually represent songs and/or collections persists, targeted either at the general public or at musically trained individuals. In this paper we describe a visual interface to browse music collections that relies on a graphical metaphor designed to convey the underlying musical structure of a song. An iconic representation of individual songs is coupled with a spatial placement of songs that reflects their structural similarity. The song icon is derived from features extracted from MIDI files, rather than from audio signals. The very nature of MIDI descriptions enables the identification of simple, yet meaningful, musical structures, allowing us to extract features that support both creating the icon and comparing songs. A similarity-based spatial placement is created projecting the feature vectors with the Least Square Projection multidimensional projection, employing the Dynamic Time Warping distance function to evaluate feature similarity. We describe the process of generating such visual representations and illustrate potentially interesting usage scenarios.