Matthias Ring

An approximation of the Gaussian RBF kernel for efficient classification with SVMs

Gaussian RBF kernels are approximated to speed up SVM classifications.An upper bound for the relative approximation error is given.Error decreases with factorial growth if approximation quality is linearly increased.Experiments showed an average 18-fold speed-up without losing accuracy. In theory, kernel support vector machines (SVMs) can be reformulated to linear SVMs. This reformulation can speed up SVM classifications considerably, in particular, if the number of support vectors is high. For the widely-used Gaussian radial basis function (RBF) kernel, however, this theoretical fact is impracticable because the reproducing kernel Hilbert space (RKHS) of this kernel has infinite dimensionality. Therefore, we derive a finite-dimensional approximative feature map, based on an orthonormal basis of the kernels RKHS, to enable the reformulation of Gaussian RBF SVMs to linear SVMs. We show that the error of this approximative feature map decreases with factorial growth if the approximation quality is linearly increased. Experimental evaluations demonstrated that the approximative feature map achieves considerable speed-ups (about 18-fold on average), mostly without losing classification accuracy. Therefore, the proposed approximative feature map provides an efficient SVM evaluation method with minimal loss of precision.

Approaching the accuracy---cost conflict in embedded classification system design

Smart embedded systems often run sophisticated pattern recognition algorithms and are found in many areas like automotive, sports and medicine. The developer of such a system is often confronted with the accuracy–cost conflict as the resulting system should be as accurate as possible while being able to run on resource constraint hardware. This article introduces a method to support the solution of this design conflict with accuracy–cost reports. These reports compare classification systems regarding their classification rate (accuracy) and the mathematical operations and parameters of the working phase (cost). Our method is used to deduce the specific cost of various popular pattern recognition algorithms and to derive the overall cost of a classification system. We also show how our analysis can be used to estimate the computational cost for specific hardware architectures. A software toolbox to create accuracy–cost reports was implemented to facilitate the automatic classification system comparison with the presented methodology. The software is available for download and as supplementary material. We performed different experiments on synthetic and real-world data to underline the value of this analysis. Accurate and computationally cheap classification systems were easily identified. We were even able to find a better implementation candidate in an existing embedded classification problem. This work is the first step towards a comprehensive support tool for the design of embedded classification systems.

Effects of stimulation frequency, amplitude, and impulse width on muscle fatigue.

Introduction: We investigated the effect of stimulation intensity (in percent of maximal tolerated stimulation current, mTSC), frequency, and impulse width on muscle fatigue. Methods: Using a randomized crossover design, 6 parameter combinations (80% mTSC, 80 Hz, 400 μs; 60% mTSC, 80 Hz, 400 μs; 80% mTSC, 20 Hz, 400 μs; 60% mTSC, 20 Hz, 400 μs; 80% mTSC, 80 Hz, 150 μs; 60% mTSC, 80 Hz, 150 μs) were tested in both legs of 13 athletic men (age 26 ± 2.3). The slope of the linear regression line over all tetani (FIS) and the number of tetani whose force was above 50% of the initial tetanus (FIN) were used to quantify fatigue. Results: FIS and FIN were significantly lower in high‐frequency protocols. No effects on FIS and FIN were found for intensity and impulse width. Conclusions: Stimulation frequency, but not impulse width or intensity, affected fatigue kinetics. Muscle Nerve 53: 608–616, 2016

Data Mining in the U.S. National Toxicology Program (NTP) Database Reveals a Potential Bias Regarding Liver Tumors in Rodents Irrespective of the Test Agent

Long-term studies in rodents are the benchmark method to assess carcinogenicity of single substances, mixtures, and multi-compounds. In such a study, mice and rats are exposed to a test agent at different dose levels for a period of two years and the incidence of neoplastic lesions is observed. However, this two-year study is also expensive, time-consuming, and burdensome to the experimental animals. Consequently, various alternatives have been proposed in the literature to assess carcinogenicity on basis of short-term studies. In this paper, we investigated if effects on the rodents’ liver weight in short-term studies can be exploited to predict the incidence of liver tumors in long-term studies. A set of 138 paired short- and long-term studies was compiled from the database of the U.S. National Toxicology Program (NTP), more precisely, from (long-term) two-year carcinogenicity studies and their preceding (short-term) dose finding studies. In this set, data mining methods revealed patterns that can predict the incidence of liver tumors with accuracies of over 80%. However, the results simultaneously indicated a potential bias regarding liver tumors in two-year NTP studies. The incidence of liver tumors does not only depend on the test agent but also on other confounding factors in the study design, e.g., species, sex, type of substance. We recommend considering this bias if the hazard or risk of a test agent is assessed on basis of a NTP carcinogenicity study.

Privacy Implications of Room Climate Data

Smart heating applications promise to increase energy efficiency and comfort by collecting and processing room climate data. While it has been suspected that the sensed data may leak crucial personal information about the occupants, this belief has up until now not been supported by evidence.

Automatic clustering of code changes

Several research tools and projects require groups of similar code changes asinput. Examples are recommendation and bug finding tools that can providevaluable information to developers based on such data. With the help ofsimilar code changes they can simplify the application of bug fixes and codechanges to multiple locations in a project. But despite their benefit, thepractical value of existing tools is limited, as users need to manually specifythe input data, i.e., the groups of similar code changes.To overcome this drawback, this paper presents and evaluates two syntacticalsimilarity metrics, one of them is specifically designed to run fast, incombination with two carefully selected and self-tuning clustering algorithmsto automatically detect groups of similar code changes.We evaluate the combinations of metrics and clustering algorithms by applyingthem to several open source projects and also publish the detected groups ofsimilar code changes online as a reference dataset. The automatically detectedgroups of similar code changes work well when used as input for LASE, arecommendation system for code changes.

Optimal feature selection for nonlinear data using branch-and-bound in kernel space

Branch-and-bound feature selection is performed in kernel space.Optimal feature subsets are found without the enormous effort of an exhaustive search.Classification accuracy is competitive to the popular wrapper approach. Branch-and-bound (B&B) feature selection finds optimal feature subsets without performing an exhaustive search. However, the classification accuracy achievable with optimal B&B feature subsets is often inferior compared to the accuracy achievable with other algorithms that guarantee optimality. We argue this is due to the existing criterion functions that define the optimal feature subset but may not conceive inherent nonlinear data structures. Therefore, we propose B&B feature selection in Reproducing Kernel Hilbert Space?(B&B-RKHS). This algorithm employs two existing criterion functions (Bhattacharyya distance, Kullback-Leibler divergence) and one new criterion function (mean class distance), however, all computed in RKHS. This enables B&B-RKHS to conceive inherent nonlinear data structures. The algorithm was experimentally compared to the popular wrapper approach that has to use an exhaustive search to guarantee optimality. The classification accuracy achieved with both methods was comparable. However, runtime of B&B-RKHS was superior using the two existing criterion functions and even completely out of reach using the new criterion function (about 60 times faster on average). Therefore, this paper proposes an efficient algorithm if feature subsets that guarantee optimality have to be selected in data sets with inherent nonlinear structures.

A Temperature-Based Bioimpedance Correction for Water Loss Estimation During Sports

The amount of total body water (TBW) can be estimated based on bioimpedance measurements of the human body. In sports, TBW estimations are of importance because mild water losses can impair muscular strength and aerobic endurance. Severe water losses can even be life threatening. TBW estimations based on bioimpedance, however, fail during sports because the increased body temperature corrupts bioimpedance measurements. Therefore, this paper proposes a machine learning method that eliminates the effects of increased temperature on bioimpedance and, consequently, reveals the changes in bioimpedance that are due to TBW loss. This is facilitated by utilizing changes in skin and core temperature. The method was evaluated in a study in which bioimpedance, temperature, and TBW loss were recorded every 15 min during a 2-h running workout. The evaluation demonstrated that the proposed method is able to reduce the error of TBW loss estimation by up to 71%, compared to the state of art. In the future, the proposed method in combination with portable bioimpedance devices might facilitate the development of wearable systems for continuous and noninvasive TBW loss monitoring during sports.

Salivary Markers for Quantitative Dehydration Estimation During Physical Exercise

Salivary markers have been proposed as noninvasive and easy-to-collect indicators of dehydrations during physical exercise. It has been demonstrated that threshold-based classifications can distinguish dehydrated from euhydrated subjects. However, considerable challenges were reported simultaneously, for example, high intersubject variabilities in these markers. Therefore, we propose a machine-learning approach to handle the intersubject variabilities and to advance from binary classifications to quantitative estimations of total body water (TBW) loss. For this purpose, salivary samples and reference values of TBW loss were collected from ten subjects during a 2-h running workout without fluid intake. The salivary samples were analyzed for previously investigated markers (osmolality, proteins) as well as additional unexplored markers (amylase, chloride, cortisol, cortisone, and potassium). Processing all these markers with a Gaussian process approach showed that quantitative TBW loss estimations are possible within an error of 0.34 l, roughly speaking, a glass of water. Furthermore, a data analysis illustrated that the salivary markers grow nonlinearly during progressive dehydration, which is in contrast to previously reported linear observations. This insight could help to develop more accurate physiological models for salivary markers and TBW loss. Such models, in turn, could facilitate even more precise TBW loss estimations in the future.

On sweat analysis for quantitative estimation of dehydration during physical exercise

Quantitative estimation of water loss during physical exercise is of importance because dehydration can impair both muscular strength and aerobic endurance. A physiological indicator for deficit of total body water (TBW) might be the concentration of electrolytes in sweat. It has been shown that concentrations differ after physical exercise depending on whether water loss was replaced by fluid intake or not. However, to the best of our knowledge, this fact has not been examined for its potential to quantitatively estimate TBW loss. Therefore, we conducted a study in which sweat samples were collected continuously during two hours of physical exercise without fluid intake. A statistical analysis of these sweat samples revealed significant correlations between chloride concentration in sweat and TBW loss (r = 0.41, p <; 0.01), and between sweat osmolality and TBW loss (r = 0.43, p <; 0.01). A quantitative estimation of TBW loss resulted in a mean absolute error of 0.49 l per estimation. Although the precision has to be improved for practical applications, the present results suggest that TBW loss estimation could be realizable using sweat samples.