Klaus Friedrichs

MOI-MBO: Multiobjective Infill for Parallel Model-Based Optimization

The aim of this work is to compare different approaches for parallelization in model-based optimization. As another alternative aside from the existing methods, we propose using a multi-objective infill criterion that rewards both the diversity and the expected improvement of the proposed points. This criterion can be applied more universally than the existing ones because it has less requirements. Internally, an evolutionary algorithm is used to optimize this criterion. We verify the usefulness of the approach on a large set of established benchmark problems for black-box optimization. The experiments indicate that the new method’s performance is competitive with other batch techniques and single-step EGO.

Tone onset detection using an auditory model

Onset detection is an important step for music transcription and other tasks frequently encountered in music processing. Although several approaches have been developed for this task, neither of them works well under all circumstances. In Bauer et al. (Einfluss der Musikinstrumente auf die Gute der Einsatzzeiterkennung, 2012) we investigated the influence of several factors like instrumentation on the accuracy of onset detection. In this work, this investigation is extended by a computational model of the human auditory periphery. Instead of the original signal the output of the simulated auditory nerve fibers is used. The main challenge here is combining the outputs of all auditory nerve fibers to one feature for onset detection. Different approaches are presented and compared. Our investigation shows that using the auditory model output leads to essential improvements of the onset detection rate for some instruments compared to previous results.

Fast Model Based Optimization of Tone Onset Detection by Instance Sampling

There exist several algorithms for tone onset detection, but finding the best one is a challenging task, as there are many categorical and numerical parameters to optimize. The aim of this task is to detect as many true onsets as possible while avoiding false detections. In recent years, model-based optimization (MBO) has been introduced for solving similar problems. The main idea of MBO is modeling the relationship between parameter settings and the response by a so-called surrogate model. After evaluating the points of an initial design—each point represents here one possible algorithm configuration—the main idea is a loop of two steps: firstly, updating a surrogate model, and secondly, proposing a new promising point for evaluation. While originally this technique has been developed mainly for numerical parameters, here, it needs to be adapted for optimizing categorical parameters as well. Unfortunately, optimization steps are very time-consuming, since the evaluation of each new point has to be performed on a large data set of music instances for getting realistic results. Nevertheless, many bad configurations could be rejected much faster, since their expected performance might appear to be very low after evaluating them on just a small partition of instances. Hence, the basic idea is to evaluate each proposed point on a small sample and only evaluate on the whole data set if the results seem to be promising.

Auralization of Auditory Models

Computational auditory models describe the transformation from acoustic signals into spike firing rates of the auditory nerves by emulating the signal transductions of the human auditory periphery.The inverse approach is called auralization, which can be useful for many tasks, such as quality measuring of signal transformations or reconstructing the hearing of impaired listeners. There have been few successful attempts to auditory inversion each of which deal with relatively simple auditory models.In recent years more comprehensive auditory models have been developed which simulate nonlinear effects in the human auditory periphery. Since for this kind of models an analytical inversion is not possible, we propose an auralization approach using statistical methods.

Comparing timbre estimation using auditory models with and without hearing loss

We propose a concept for evaluating signal transformations for music signals with respect to an individual hearing deficit by using an auditory model. This deficit is simulated in the model by changing specific model parameters. Our idea is extracting the musical attributes rhythm, pitch, loudness and timbre and comparing the modified model output to the original one. While rhythm, pitch, and loudness estimation are studied in previous works the focus in this paper concentrates on timbre estimation. Results are shown for the original auditory model and three models, each simulating a specific hearing loss.

A multivariate approach for onset detection using supervised classification

In this paper we introduce a new onset detection approach which incorporates a supervised classification model for estimating the tone onset probability in signal frames. In contrast to the most classical strategies where only one detection function can be applied for signal feature extraction, the classification model can be fitted on a large feature set. This is meaningful since, depending on the music characteristics, some detection functions can be more advantageous that the others. Although the idea of the considering of many detection functions is not new in the literature, these functions are, so far, treated in a univariate way by, e.g., building of weighted sums. This probably lies on the difficulties of the direct transfer of the classification ideas to the onset detection task. The goodness measure of onset detection is namely based on the comparison of two time vectors while by the classification such a measure is derived from the framewise matches of predicted and true labels. In this work we first construct – based on several resent publications – a comprehensive univariate onset detection algorithm which depends on many free settable parameters. Then, the new multivariate approach also depending on many free parameters is introduced. The parameters of the both onset detection strategies are optimized for online and offline cases by utilizing an appropriate validation technique. The main funding is that the multivariate strategy outperforms the univariate one significantly regarding the F -measure. Furthermore, the multivariate approach seems to be especially beneficial in online case since it requires only the halve of the future signal information comparing to the best setting of the univariate onset detection.

Time efficient optimization of instance based problems with application to tone onset detection

A time efficient optimization technique for instance based problems is proposed, where for each parameter setting the target function has to be evaluated on a large set of problem instances. Computational time is reduced by beginning with a performance estimation based on the evaluation of a representative subset of instances. Subsequently, only promising settings are evaluated on the whole data set. As application a comprehensive music onset detection algorithm is introduced where several numerical and categorical algorithm parameters are optimized simultaneously. Here, problem instances are music pieces of a data base. Sequential model based optimization is an appropriate technique to solve this optimization problem. The proposed optimization strategy is compared to the usual model based approach with respect to the goodness measure for tone onset detection. The performance of the proposed method appears to be competitive with the usual one while saving more than 84% of instance evaluation time on average. One other aspect is a comparison of two strategies for handling categorical parameters in Kriging based optimization.

Efficient global optimization: Motivation, variations and applications

A popular optimization method of a black box objective function is Efficient Global Optimization (EGO), also known as Sequential Model Based Optimization, SMBO, with kriging and expected improvement. EGO is a sequential design of experiments aiming at gaining as much information as possible from as few experiments as feasible by a skillful choice of the factor settings in a sequential way. In this paper we will introduce the standard procedure and some of its variants. In particular, we will propose some new variants like regression as a modeling alternative to kriging and two simple methods for the handling of categorical variables, and we will discuss focus search for the optimization of the infill criterion. Finally, we will give relevant examples for the application of the method. Moreover, in our group, we implemented all the described methods in the publicly available R package mlrMBO.

Model based optimization of music onset detection

In this paper a comprehensive online music onset detection algorithm is introduced where – in contrast to many other relevant publications – 14 important algorithm parameters are optimized simultaneously. For solving the optimization problem we derive an extensive tool for iterative model based optimization. In each iteration, a very time consuming evaluation has to be performed on a large music data base. To speed up this procedure, the expected performance of each newly proposed setting is estimated in a pretest on a representative part of the data so that just very promising points are evaluated on all data. We compare different variants of the classical and the fast optimization strategies with respect to the F -values of their best identified parameter settings. The performance of the fast approach appears to be competitive with the classical one while saving more than 80% of music piece evaluations on average. Our best found parameter settings, both for online and offline onset detection, are mainly in accordance with the usual choices in the stateof-the art literature concerning, e.g., the spectral flux detection function or preferences for window length and overlap. However, we also found unexpected results. For example, the adaptive whitening pre-processing step showed no effect.

Software in Music Information Retrieval

Music Information Retrieval (MIR) software is often applied for the identification of rules classifying audio music pieces into certain categories, like e.g. genres. In this paper we compare the abilities of six MIR software packages in ten categories, namely operating systems, user interface, music data input, feature generation, feature formats, transformations and features, data analysis methods, visualization methods, evaluation methods, and further development. The overall rankings are derived from the estimated scores for the analyzed criteria.