1 - 10 of 15
-
PublicationExplainable AI for sensor-based sorting systems( 2023)Explainable artificial intelligence (XAI) can make machine learning based systems more transparent. This additional transparency can enable the use of machine learning in many different domains. In our work, we show how XAI methods can be applied to an autoencoder for anomaly detection in a sensor-based sorting system. The setup of the sorting system consists of a vibrating feeder, a conveyor belt, a line-scan camera and an array of fast-switching pneumatic valves. It allows the separation of a material stream into two fractions, realizing a binary sorting task. The autoencoder tries to mimic the normal behavior of the nozzle array and thus can detect abnormal behavior. The XAI methods are used to explain the output of the autoencoder. As XAI methods global and local approaches are used, which means we receive explanations for both a single result and the whole autoencoder. Initial results for both approaches are shown, together with possible interpretations of these results
-
PublicationValidation of XAI Explanations for Multivariate Time Series Classification in the Maritime Domain( 2022)Due to the lack of explanation towards their internal mechanism, state-of-the-art deep learning-based classifiers are often considered as black-box models. For instance, in the maritime domain, models that classify the types of ships based on their trajectories and other features perform well, but give no further explanation for their predictions. To gain the trust of human operators responsible for critical decisions, the reason behind the classification is crucial. In this paper, we introduce explainable artificial intelligence (XAI) approaches to the task of classification of ship types. This supports decision-making by providing explanations in terms of the features contributing the most towards the prediction, along with their corresponding time intervals. In the case of the LIME explainer, we adapt the time-slice mapping technique (LimeforTime), while for Shapley additive explanations (SHAP) and path integrated gradient (PIG), we represent the relevance of each input variable to generate a heatmap as an explanation. In order to validate the XAI results, the existing perturbation and sequence analyses for classifiers of univariate time series data is employed for testing and evaluating the XAI explanations on multivariate time series. Furthermore, we introduce a novel evaluation technique to assess the quality of explanations yielded by the chosen XAI method.
-
PublicationA Survey on the Explainability of Supervised Machine Learning( 2021)Predictions obtained by, e.g., artificial neural networks have a high accuracy but humans often perceive the models as black boxes. Insights about the decision making are mostly opaque for humans. Particularly understanding the decision making in highly sensitive areas such as healthcare or finance, is of paramount importance. The decision-making behind the black boxes requires it to be more transparent, accountable, and understandable for humans. This survey paper provides essential definitions, an overview of the different principles and methodologies of explainable Supervised Machine Learning (SML). We conduct a state-of-the-art survey that reviews past and recent explainable SML approaches and classifies them according to the introduced definitions. Finally, we illustrate principles by means of an explanatory case study and discuss important future directions.
-
PublicationSupported Decision-Making by Explainable Predictions of Ship Trajectories( 2021)Machine Learning and Deep Learning models make accurate predictions based on a specifically trained task. For instance, models that classify ship vessel types based on their trajectory and other features. This can support human experts while they try to obtain information on the ships, e.g., to control illegal fishing. Besides the support in predicting a certain ship type, there is a need to explain the decision-making behind the classification. For example, which features contributed the most to the classification of the ship type. This paper introduces existing explanation approaches to the task of ship classification. The underlying model is based on a Residual Neural Network. The model was trained on an AIS data set. Further, we illustrate the explainability approaches by means of an explanatory case study and conduct a first experiment with a human expert.
-
PublicationASARob - Aufmerksamkeitssensitiver Assistenzroboter( 2020)Das Projektziel des Vorhabens ASARob war die Implementierung einer robusten Aufmerksamkeitserfassung und -lenkung für die Roboter-Mensch-Interaktion. Multimodale Verfahren zur Aufmerksamkeitserfassung und -lenkung wurden hierzu in die bestehende, mobile Roboterplattform Care-O-bot 4 (care-o-bot.de) integriert. Die fusionierten Verfahren dienten als zentrale Grundfertigkeiten des Roboters, um bestehende Assistenzfunktionen, wie z. B. das räumliche Führen zu vorgegebenen Orten oder das Holen und Bringen von Gegenständen, anzureichern und in einem intuitiven Dialog mit dem betroffenen Nutzer durchführen zu können. Die Aufmerksamkeitserfassung diente insbesondere zur erwartungskonformen und kontextangepassten Annäherung des Roboters an Menschen bzw. Gesprächspartner. Durch den Einsatz der multimodalen Erfassungsvielfalt von Menschen und Umfeld sollte insbesondere gewährleistet werden, dass die Aufmerksamkeit von Personen auch in unstrukturierten Umgebungen, wie sie im Alltag zu erwarten sind, robust und fehlertolerant nachvollzogen werden kann. So wurden bspw. durch die Erfassung der Blickrichtung, Kopfdrehung, Sprache, Stimme und Körperhaltung von Nutzern partiell redundante Wahrnehmungskanäle implementiert, die sich gegenseitig ergänzen, insbesondere aber bei etwaigem Ausfall eines Kanals (z. B. durch Hinterkopfansichten, die eine Sicht auf die Augen eines Nutzers verhindern) durch konfidenzbasierte Informationsfusion für Rückfalloptionen sorgen.
-
PublicationDeutsche Normungsroadmap Künstliche Intelligenz(DIN, 2020)
-
PublicationBatch-wise Regularization of Deep Neural Networks for Interpretability( 2020)Fast progress in the field of Machine Learning and Deep Learning strongly influences the research in many application domains like autonomous driving or health care. In this paper, we propose a batch-wise regularization technique to enhance the interpretability for deep neural networks (NN) by means of a global surrogate rule list. For this purpose, we introduce a novel regularization approach that yields a differentiable penalty term. Compared to other regularization approaches, our approach avoids repeated creating of surrogate models during training of the NN. The experiments show that the proposed approach has a high fidelity to the main model and also results in interpretable and more accurate models compared to some of the baselines.
-
PublicationCognitive Systems and Robotics( 2019)Cognitive systems are able to monitor and analyze complex processes, which also provides them with the ability to make the right decisions in unplanned or unfamiliar situations. Fraunhofer experts are employing machine learning techniques to harness new cognitive functions for robots and automation solutions. To do this, they are equipping systems with technologies that are inspired by human abilities, or imitate and optimize them. This report describes these technologies, illustrates current example applications, and lays out scenarios for future areas of application.
-
PublicationGaussian Process based Dynamic Facial Emotion Tracking( 2019)Capturing the emotions of humans is of paramount importance in human-machine interaction. Here, emotions are typically extracted from the human's face recorded in image sequences. In this paper, tracking emotions from images is formulated as Bayesian state estimation problem where the system state represents the valence-arousal space of emotions. Handcrafted image features are first mapped to the valence-arousal space by means of a Gaussian process. To allow dynamic emotion tracking, a Kalman filter is derived, where an inequality constraint on the emotional state is employed in order to avoid a drifting state. Experiments based on two well-known facial expression datasets are performed to demonstrate the performance of the proposed approach.
-
PublicationComparison of Angle and Size Features with Deep Learning for Emotion Recognition( 2019)The robust recognition of a person's emotion from images is an important task in human-machine interaction. This task can be considered a classification problem, for which a plethora of methods exists. In this paper, the emotion recognition performance of two fundamentally different approaches is compared: classification based on hand-crafted features against deep learning. This comparison is conducted by means of well-established datasets and highlights the benefits and drawbacks of each approach.
