Borgelt, Christian
E-Mail-Adresse
ORCID
Geburtsdatum
Berufsbeschreibung
Nachname
Vorname
Name
Suchergebnisse Publikationen
Differentiable Top-k Classification Learning
The top-k classification accuracy is one of the core metrics in machine learning. Here, k is conventionally a positive integer, such as 1 or 5, leading to top-1 or top-5 training objectives. In this work, we relax this assumption and optimize the model for multiple k simultaneously instead of using a single k. Leveraging recent advances in differentiable sorting and ranking, we propose a family of differentiable top-k cross-entropy classification losses. This allows training while not only considering the top-1 prediction, but also, e.g., the top-2 and top-5 predictions. We evaluate the proposed losses for fine-tuning on state-of-the-art architectures, as well as for training from scratch. We find that relaxing k not only produces better top-5 accuracies, but also leads to top-1 accuracy improvements. When fine-tuning publicly available ImageNet models, we achieve a new state-of-the-art for these models.
Even Faster Exact k-Means Clustering
A naïve implementation of k-means clustering requires computing for each of the n data points the distance to each of the k cluster centers, which can result in fairly slow execution. However, by storing distance information obtained by earlier computations as well as information about distances between cluster centers, the triangle inequality can be exploited in different ways to reduce the number of needed distance computations, e.g. [3, 4, 5, 7, 11]. In this paper I present an improvement of the Exponion method [11] that generally accelerates the computations. Furthermore, by evaluating several methods on a fairly wide range of artificial data sets, I derive a kind of map, for which data set parameters which method (often) yields the lowest execution times.
Fuzzy Sets in Data Analysis : From Statistical Foundations to Machine Learning
Basic ideas and formal concepts from fuzzy sets and fuzzy logic have been used successfully in various branches of science and engineering. This paper elaborates on the use of fuzzy sets in the broad field of data analysis and statistical sciences, including modern manifestations such as data mining and machine learning. In the fuzzy logic community, this branch of research has recently gained in importance, especially due to the emergence of data science as a new scientific discipline, and the increasing relevance of machine learning as a key methodology of modern artificial intelligence. This development has been accompanied by an internal shift from largely knowledge-based to strongly data-driven fuzzy modeling and systems design. Reflecting on the historical dimension and evolution of the area, we discuss the role of fuzzy logic in data analysis and related fields, highlight existing contributions of fuzzy sets in these fields, and outline interesting directions for future work.
Learned Feature Generation for Molecules
When classifying molecules for virtual screening, the molecular structure first needs to be converted into meaningful features, before a classifier can be trained. The most common methods use a static algorithm that has been created based on domain knowledge to perform this generation of features. We propose an approach where this conversion is learned by convolutional neural network finding features that are useful for teh task at hand based on the available data. Preliminary results indicate that our current approach can already come up with fetaures that perform similarly well as common methods. Since this approach does not jet use any chemiocal properties, results could be improved in future versions
GenDR : A Generalized Differentiable Renderer
In this work, we present and study a generalized family of differentiable renderers. We discuss from scratch which components are necessary for differentiable rendering and formalize the requirements for each component. We instantiate our general differentiable renderer, which generalizes existing differentiable renderers like SoftRas and DIB-R, with an array of different smoothing distributions to cover a large spectrum of reasonable settings. We evaluate an array of differentiable renderer instantiations on the popular ShapeNet 3D reconstruction benchmark and analyze the implications of our results. Surprisingly, the simple uniform distribution yields the best overall results when averaged over 13 classes; in general, however, the optimal choice of distribution heavily depends on the task.
Guide to Intelligent Data Science : How to Intelligently Make Use of Real Data
Making use of data is not anymore a niche project but central to almost every project. With access to massive compute resources and vast amounts of data, it seems at least in principle possible to solve any problem. However, successful data science projects result from the intelligent application of: human intuition in combination with computational power; sound background knowledge with computer-aided modelling; and critical reflection of the obtained insights and results. Substantially updating the previous edition, then entitled Guide to Intelligent Data Analysis, this core textbook continues to provide a hands-on instructional approach to many data science techniques, and explains how these are used to solve real world problems. The work balances the practical aspects of applying and using data science techniques with the theoretical and algorithmic underpinnings from mathematics and statistics. Major updates on techniques and subject coverage (including deep learning) are included.
Aggregation of Subclassifications : Methods, Tools and Experiments
Aggregation methods have been studied extensively from a mathematical, theoretical point of view. In this work, however, we focus on a more practical aspect: subclassifications. Given class predictions for several sub-objects of a single instance, we systematically investigate the performance of different aggregation methods. To this end, we simulate data for various data distributions. Thus we ensure that we know the ground truth for the evaluation, which would be impossible for real world data. Our source code is publicly available and can be extended to explore other aggregation methods and other data distributions.
Differentiable Sorting Networks for Scalable Sorting and Ranking Supervision
Sorting and ranking supervision is a method for training neural networks end-to-end based on ordering constraints. That is, the ground truth order of sets of samples is known, while their absolute values remain unsupervised. For that, we propose differentiable sorting networks by relaxing their pairwise conditional swap operations. To address the problems of vanishing gradients and extensive blurring that arise with larger numbers of layers, we propose mapping activations to regions with moderate gradients. We consider odd-even as well as bitonic sorting networks, which outperform existing relaxations of the sorting operation. We show that bitonic sorting networks can achieve stable training on large input sets of up to 1024 elements.
Widened Learning of Index Tracking Portfolios
Index investing has an advantage over active investment strategies, because less frequent trading results in lower expenses, yielding higher long-term returns. Index tracking is a popular investment strategy that attempts to find a portfolio replicating the performance of a collection of investment vehicles. This paper considers index tracking from the perspective of solution space exploration. Three search space heuristics in combination with three portfolio tracking error methods are compared in order to select a tracking portfolio with returns that mimic a benchmark index. Experimental results conducted on real-world datasets show that Widening, a metaheuristic using diverse parallel search paths, finds superior solutions than those found by the reference heuristics. Presented here are the first results using Widening on time-series data.
Training Neural Networks to Distinguish Craving Smokers, Non-craving Smokers, and Non-smokers
In the present study, we investigate the differences in brain signals of craving smokers, non-craving smokers, and non-smokers. To this end, we use data from resting-state EEG measurements to train predictive models to distinguish these three groups. We compare the results obtained from three simple models – majority class prediction, random guessing, and naive Bayes – as well as two neural network approaches. The first of these approaches uses a channel-wise model with dense layers, the second one uses cross-channel convolution. We therefore generate a benchmark on the given data set and show that there is a significant difference in the EEG signals of smokers and non-smokers.