Search

This study proposes a framework for the automated hyperparameter optimization of a bearing fault detection pipeline for permanent magnet synchronous motors (PMSMs) without the need of external sensors. A automated machine learning (AutoML) pipeline search is performed by means of a genetic optimization to reduce human induced bias due to inappropriate parameterizations. For this purpose, a search space is defined, which includes general methods of signal processing and manipulation as well as methods tailored to the respective task and domain. The proposed framework is evaluated on the bearing fault detection use case under real world conditions. Considerations on the generalization of the deployed fault detection pipelines are also taken into account. Likewise, attention was paid to experimental studies for evaluations of the robustness of the fault detection pipeline to variations of the motors working condition parameters between the training and test domain. The present work contributes to the research of fault detection on rotating machinery in the following terms: (1) Reduction of the human induced bias to the data science process, while still considering expert and task related knowledge, ending in a generic search approach (2) tackling the bearing fault detection task without the need for external sensors (sensorless) (3) learning a domain robust fault detection pipeline applicable to varying motor operating parameters without the need of re-parameterizations or fine-tuning (4) investigations on working condition discrepancies with an excessive degree to determine the pipeline limitations regarding the abstraction of the motor parameters and the pipeline hyperparameters, 8 pages, 4 figures, 5 tables, ieee conference paper template used

We present an empirical study on the use of continual learning (CL) methods in a reinforcement learning (RL) scenario, which, to the best of our knowledge, has not been described before. CL is a very active recent research topic concerned with machine learning under non-stationary data distributions. Although this naturally applies to RL, the use of dedicated CL methods is still uncommon. This may be due to the fact that CL methods often assume a decomposition of CL problems into disjoint sub-tasks of stationary distribution, that the onset of these sub-tasks is known, and that sub-tasks are non-contradictory. In this study, we perform an empirical comparison of selected CL methods in a RL problem where a physically simulated robot must follow a racetrack by vision. In order to make CL methods applicable, we restrict the RL setting and introduce non-conflicting subtasks of known onset, which are however not disjoint and whose distribution, from the learner's point of view, is still non-stationary. Our results show that dedicated CL methods can significantly improve learning when compared to the baseline technique of "experience replay"., Comment: Accepted at the IJCNN2022, 9 pages, 9 figures

Continual Learning (CL, sometimes also termed incremental learning) is a flavor of machine learning where the usual assumption of stationary data distribution is relaxed or omitted. When naively applying, e.g., DNNs in CL problems, changes in the data distribution can cause the so-called catastrophic forgetting (CF) effect: an abrupt loss of previous knowledge. Although many significant contributions to enabling CL have been made in recent years, most works address supervised (classification) problems. This article reviews literature that study CL in other settings, such as learning with reduced supervision, fully unsupervised learning, and reinforcement learning. Besides proposing a simple schema for classifying CL approaches w.r.t. their level of autonomy and supervision, we discuss the specific challenges associated with each setting and the potential contributions to the field of CL in general., Comment: Accepted at the ESANN2022, 19 pages, 1 figure

We present an approach for continual learning (CL) that is based on fully probabilistic (or generative) models of machine learning. In contrast to, e.g., GANs that are "generative" in the sense that they can generate samples, fully probabilistic models aim at modeling the data distribution directly. Consequently, they provide functionalities that are highly relevant for continual learning, such as density estimation (outlier detection) and sample generation. As a concrete realization of generative continual learning, we propose Gaussian Mixture Replay (GMR). GMR is a pseudo-rehearsal approach using a Gaussian Mixture Model (GMM) instance for both generator and classifier functionalities. Relying on the MNIST, FashionMNIST and Devanagari benchmarks, we first demonstrate unsupervised task boundary detection by GMM density estimation, which we also use to reject untypical generated samples. In addition, we show that GMR is capable of class-conditional sampling in the way of a cGAN. Lastly, we verify that GMR, despite its simple structure, achieves state-of-the-art performance on common class-incremental learning problems at very competitive time and memory complexity., Accepted as Findings at the CLVISION2021 workshop, 11 pages, 6 figures

Research on computer vision systems for driver assistance resulted in a variety of isolated approaches mainly performing very specialized tasks like, e. g., lane keeping or traffic sign detection. However, for a full understanding of generic traffic situations, integrated and flexible approaches are needed. We here present a highly integrated vision architecture for an advanced driver assistance system inspired by human cognitive principles. The system uses an attention system as the flexible and generic front-end for all visual processing, allowing a task-specific scene decomposition and search for known objects (based on a short term memory) as well as generic object classes (based on a long term memory). Knowledge fusion, e. g., between an internal 3D representation and a reliable road detection module improves the system performance. The system heavily relies on top-down links to modulate lower processing levels, resulting in a high system robustness.

This work presents a mathematical treatment of the relation between Self-Organizing Maps (SOMs) and Gaussian Mixture Models (GMMs). We show that energy-based SOM models can be interpreted as performing gradient descent, minimizing an approximation to the GMM log-likelihood that is particularly valid for high data dimensionalities. The SOM-like decrease of the neighborhood radius can be understood as an annealing procedure ensuring that gradient descent does not get stuck in undesirable local minima. This link allows to treat SOMs as generative probabilistic models, giving a formal justification for using SOMs, e.g., to detect outliers, or for sampling., Comment: 10 pages, 2 figures, submitted and accepted at International Conference on Artificial Neural Networks (ICANN) 2020

We present a study of deep learning applied to the domain of network traffic data forecasting. This is a very important ingredient for network traffic engineering, e.g., intelligent routing, which can optimize network performance, especially in large networks. In a nutshell, we wish to predict, in advance, the bit rate for a transmission, based on low-dimensional connection metadata ("flows") that is available whenever a communication is initiated. Our study has several genuinely new points: First, it is performed on a large dataset (~50 million flows), which requires a new training scheme that operates on successive blocks of data since the whole dataset is too large for in-memory processing. Additionally, we are the first to propose and perform a more fine-grained prediction that distinguishes between low, medium and high bit rates instead of just "mice" and "elephant" flows. Lastly, we apply state-of-the-art visualization and clustering techniques to flow data and show that visualizations are insightful despite the heterogeneous and non-metric nature of the data. We developed a processing pipeline to handle the highly non-trivial acquisition process and allow for proper data preprocessing to be able to apply DNNs to network traffic data. We conduct DNN hyper-parameter optimization as well as feature selection experiments, which clearly show that fine-grained network traffic forecasting is feasible, and that domain-dependent data enrichment and augmentation strategies can improve results. An outlook about the fundamental challenges presented by network traffic analysis (high data throughput, unbalanced and dynamic classes, changing statistics, outlier detection) concludes the article., Comment: 16 pages, 12 figures, 28th International Conference on Artificial Neural Networks (ICANN 2019)

International audience; We present a trainable hierarchical architecture capable of detecting newness (or outliers) at all hierarchical levels. This contribution paves the way for deep neural architectures that are able to learn in an incremental fashion, for which the ability to detect newness is an indispensable prerequisite. We verify the ability to detect newness by conducting experiments on the MNIST database, where we introduce either localized changes, by adding noise to a small patch of the input, or global changes, by changing the global arrangement of local patterns which is not detectable at the local level.

In this document, I first analyze some of the reasons why real-world environment perception is stillstrongly inferior to human perception in overall accuracy and reliability. In particular, I focus on thetask of object detection in traffic scenes and present an argument why this task is in fact a good modeltask for other, related perception problems (e.g. in robotics or surveillance). Enumerating the difficultiesencountered in this model task (and therefore, by inference, in many other detection tasks as well), I cometo the conclusion that problems in object detection can in fact be, to a significant extent, traced backto problems of the learning algorithms that are used in various forms when performing object detection.Namely, the lack of a probabilistic interpretation, the lack of incremental learning capacity, the lack oftraining samples and the inherent ambiguity of local pattern analysis are identified and used to justify aroad map for research efforts aimed at overcoming these problems. I present several of my works concerningreal-world applications of machine learning in perception, where the stated problems become very apparent.Subsequently, I describe in detail my recent research contributions and their significance in the context ofthe proposed road map: context-based object detection, generative and multi-modal learning as well as anoriginal method for incremental learning. The document is concluded by an outlook that addresses furtherwork to complete the road map, and the possibilities that are offered by such an endeavour in the field of machine perception.

We present a novel method to perform multi-class pattern classification with neural networks and test it on a challenging 3D hand gesture recognition problem. Our method consists of a standard one-against-all (OAA) classification, followed by another network layer classifying the resulting class scores, possibly augmented by the original raw input vector. This allows the network to disambiguate hard-to-separate classes as the distribution of class scores carries considerable information as well, and is in fact often used for assessing the confidence of a decision. We show that by this approach we are able to significantly boost our results, overall as well as for particular difficult cases, on the hard 10-class gesture classification task., European Symposium on artificial neural networks (ESANN), Jun 2015, Bruges, Belgium

International audience; This study is conducted in the context of developmental learning in embodied agents who have multiple data sources (sensors) at their disposal. We describe an online learning method that simultaneously discovers "meaningful" concepts in the associated processing streams, extending methods such as PCA, SOM or sparse coding to the multimodal case. In addition to the avoidance of redundancies in the concepts derived from single modalities, we claim that "meaningful" concepts are those who have statistical relations across modalities. This is a reasonable claim because measurements by different sensors often have common cause in the external world and therefore carry correlated information. To capture such cross-modal relations while avoiding redundancy of concepts, we propose a set of interacting self-organization processes which are modulated by local predictability. To validate the fundamental applicability of the method, we conduct a plausible simulation experiment with synthetic data and find that those concepts which are predictable from other modalities successively "grow", i.e., become overrepresented, whereas concepts that are not predictable become systematically under-represented. We conclude the article by a discussion of applicability in real-world robotics scenarios.

Research on computer vision systems for driver assistance resulted in a variety of approaches mainly performing reactive tasks like, e.g., lane keeping. However, for a full understanding of generic traffic situations, integrated and more flexible approaches are needed. We present a system inspired by the human visual system. Based on combining task-dependent tunable visual saliency, an object recognizer, and a tracker it provides warnings in dangerous situations., International Conference on Computer Vision Systems : Proceedings, The 5th International Conference on Computer Vision Systems (ICVS)

Die Arbeit verfolgt praxisorientierte sowie grundlagenorientierte Aspekte zur Optimierung von Objekterkennungssystemen durch die Weiterentwicklung evolutionärer und neuronaler Verfahren. Es wird der Einsatz neuronaler Klassifikatoren im Rahmen eines Echtzeit-Autoerkennungssystems dargestellt sowie ein systematisches Verfahren zur evolutionären Mehrziel- Strukturanpassung neuronaler Klassifikatoren entwickelt. Weiterhin wird ein lernfähiges Salienzkartenmodell präsentiert, um initiale Objektdetektion in Analogie zu Mechanismen selektiver visueller Aufmerksamkeit zu lösen. Das Modell ist anhand von Beispielen trainierbar; Wettbewerbsmechanismen sind vollständig dynamisch formuliert. Schließlich wird Echtzeit-Objekterkennung präsentiert, welche auf Ganzbildsuche mit einem neuen Typ von Faltungsnetzwerk (CNN) basiert. Ein hybrides überwacht/unüberwachtes Verfahren wird benutzt, um systematisch geeignete Merkmale sowie die korrekte Klassifikationsentscheidung aus den Bilddaten zu lernen.

Many state of the art object classification applications require many data samples, whose collection is usually a very costly process. Performing initial model training with synthetic samples (from virtual reality tools) has been proposed as a possible solution, although the resulting classification models need to be adapted (fine-tuned) to real-world data afterwards. For this bootstrapping process, we propose to use an incremental learning algorithm from the cognitive robotics domain which is particularly suited for perceptual problems. We apply it to a pedestrian detection problem where a synthetic dataset is used for initial training, and two different real-world datasets for fine-tuning and evaluation. The proposed scheme greatly reduces the number of real-world samples required while maintaining high classification accuracy. We additionally demonstrate several innovative incremental learning schemes for object detection, the basic idea being that usually only very few background samples are actually similar to pedestrian samples. By suitable arrangement of incremental learning steps, we can keep classification models simple by representing only such “hard” background samples.

International audience; In the context of sensory or higher-level cognitive processing, we present a recurrent neural network model, similar to the popular dynamic neural field (DNF) model, for performing approximate probabilistic computations. The model is biologically plausible, avoids impractical schemes such as log-encoding and noise assumptions, and is well-suited for working in stacked hierarchies. By Lyapunov analysis, we make it very plausible that the model computes the maximum a posteriori (MAP) estimate given a certain input that may be corrupted by noise. Key points of the model are its capability to learn the required posterior distributions and represent them in its lateral weights, the interpretation of stable neural activities as MAP estimates, and of latency as the probability associated with those estimates. We demonstrate for in simple experiments that learning of posterior distributions is feasible and results in correct MAP estimates. Furthermore, a pre-activation of field sites can modify attractor states when the data model is ambiguous, effectively providing an approximate implementation of Bayesian inference.

International audience; Self-organizing map (SOM) is a powerful paradigm that is extensively applied for clustering and visualization purpose. It is also used for regression learning, especially in robotics, thanks to its ability to provide a topological projection of high dimensional non linear data. In this case, data extracted from the SOM are usually restricted to the best matching unit (BMU), which is the usual way to use SOM for classification , where class labels are attached to individual neurons. In this article, we investigate the influence of considering more information from the SOM than just the BMU when performing regression. For this purpose , we quantitatively study several output functions for the SOM, when using these data as input of a linear regression, and find that the use of additional activities to the BMU can strongly improve regression performance. Thus, we propose an unified and generic framework that embraces a large spectrum of models from the traditional way to use SOM, with the best matching unit as output, to models related to the radial basis function network paradigm, when using local receptive field as output.

International audience; Building upon prior results, we present an alterna-tive approach to efficiently classifying a complex set of 3D hand poses obtained from modern Time-Of-Flight-Sensors (TOF). We demonstrate it is possible to achieve satisfactory results in spite of low resolution and high noise (inflicted by the sensors) and a demanding outdoor environment. We set up a large database of pointclouds in order to train multilayer perceptrons as well as support vector machines to classify the various hand poses. Our goal is to fuse data from multiple TOF sensors, which observe the poses from multiple angles. The presented contribution illustrates that real-time capability can be maintained with such a setup as the used 3D descriptors, the fusion strategy as well as the online confidence measures are computationally efficient.

In this contribution, we present a large-scale hierarchical system for object detection fusing bottom-up (signal-driven) processing results with top-down (model or task-driven) attentional modulation. Specifically, we focus on the question of how the autonomous learning of invariant models can be embedded into a performing system and how such models can be used to define object-specific attentional modulation signals. Our system implements bi-directional data flow in a processing hierarchy. The bottom-up data flow proceeds from a preprocessing level to the hypothesis level where object hypotheses created by exhaustive object detection algorithms are represented in a roughly retino- topic way. A competitive selection mechanism is used to determine the most confident hypotheses, which are used on the system level to train multimodal models that link object identity to invariant hypothesis properties. The top-down data flow originates at the system level, where the trained multimodal models are used to obtain space- and feature- based attentional modulation signals, providing biases for the competitive selection process at the hypothesis level. This results in object-specific hypothesis facilitation/sup- pression in certain image regions which we show to be applicable to different object detection mechanisms. In order to demonstrate the benefits of this approach, we apply the system to the detection of cars in a variety of chal- lenging traffic videos. Evaluating our approach on a pub- licly available dataset containing approximately 3,500 annotated video images from more than 1 h of driving, we can show strong increases in performance and generaliza- tion when compared to object detection in isolation. Furthermore, we compare our results to a late hypothesis rejection approach, showing that early coupling of top- down and bottom-up information is a favorable approach especially when processing resources are constrained.

International audience; Incremental learning refers to learning from streaming data, which arrive over time, with limited memory resources and, ideally, without sacrificing model accuracy. This setting fits different application scenarios where lifelong learning is relevant, e.g. due to changing environments , and it offers an elegant scheme for big data processing by means of its sequential treatment. In this contribution, we formalise the concept of incremental learning, we discuss particular challenges which arise in this setting, and we give an overview about popular approaches, its theoretical foundations, and applications which emerged in the last years.