Machine learning classifiers adjust implicitly or explicitly the importance of the data features to solve a given classification task.This feature weighting often does not imply feature sparseness, which, however, may be important for interpretability and model evaluation.This contribution proposes how to force feature sparseness in combination with feature relevance for prototype-based classification learning to obtain reliable and interpretable classification decisions.

Prototype-based classification learning methods are known to be inherently interpretable. However, this paradigm suffers from major limitations compared to deep models, such as lower performance. This led to the development of the so-called deep Prototype-Based Networks (PBNs), also known as prototypical parts models. In this work, we analyze these models with respect to different properties, including interpretability. In particular, we focus on the Classification-by-Components (CBC) approach, which uses a probabilistic model to ensure interpretability and can be used as a shallow or deep architecture. We show that this model has several shortcomings, like creating contradicting explanations. Based on these findings, we propose an extension of CBC that solves these issues. Moreover, we prove that this extension has robustness guarantees and derive a loss that optimizes robustness. Additionally, our analysis shows that most (deep) PBNs are related to (deep) RBF classifiers, which implies that our robustness guarantees generalize to shallow RBF classifiers. The empirical evaluation demonstrates that our deep PBN yields state-of-the-art classification accuracy on different benchmarks while resolving the interpretability shortcomings of other approaches. Further, our shallow PBN variant outperforms other shallow PBNs while being inherently interpretable and exhibiting provable robustness guarantees.

Knowledge integration into machine learning systems is a promising and successful strategy to achieve more plausible and consistent results.The plausibility is accompanied by better model interpretability due to the adjustment of the machine learning system to the domain specic requirements and restrictions.Further, informed machine learning can be seen as a particular task specic regularization of the model leading to better learning convergence and frequently also requiring a lower amount of training data.This short introduction paper addresses some recent aspects, how domain knowledge can be integrated into learning systems on dierent levels ranging from informed feature extraction to domain adjusted structure and model architecture.* M.K. is supported by the IAI-

Provident detection of other road users at night has the potential for increasing road safety. For this purpose, humans intuitively use visual cues, such as light cones and light reflections emitted by other road users to be able to react to oncoming traffic at an early stage. Computer vision methods can imitate this behavior by predicting the appearance of vehicles based on light reflections caused by the vehicle’s headlights. Since current object detection algorithms are mainly based on detecting directly visible objects annotated via bounding boxes, the detection and annotation of light reflections without sharp boundaries is challenging. For this reason, the extensive open-source PVDN (Provident Vehicle Detection at Night) dataset was published that includes traffic scenarios at night with light reflections annotated via keypoints. In this paper, we explore a generic approach to annotate objects without clear boundaries, such as light reflections, by combining sparse keypoint annotations of humans with the concept of Boolean map saliency. With that, we create context-aware saliency maps that capture unsharp object boundaries, such as of light reflections. We show that this approach allows for an automated derivation of different object representations, such as bounding boxes, so that detection models can be trained and the problem of providently detecting vehicles at night can be tackled from a different perspective. Our approach makes it possible to derive bounding boxes with superior quality compared to previous approaches and to develop better object detection algorithms. With this paper, we provide a powerful method to study the problem of detecting objects with unsharp boundaries and, in particular, to investigate the detection of vehicles at night before they are actually visible.

Detecting oncoming vehicles at night as early as possible is important for highly automated driving and Advanced-Driver-Assistance-Systems (ADAS). The sooner objects are detected, the earlier autonomous systems can take them into consideration to plan more anticipatory and safe actions. Previous work showed that on rural land roads at night, oncoming vehicles can already be detected before they are actually directly visible. This is done based on their emitted light. However, no work exists on covering the problem for more complex scenarios such as urban areas in cities. In this paper, we present a new approach to annotate light reflections in urban scenarios caused by oncoming vehicles at night before they are directly visible. We revisit design decisions in previous work for rural land road scenarios and find several improvements. We propose a pipeline which takes relatively cheap-to-get, yet highly subjective human Bounding-Box (BB) annotations, and automatically turns them into normally expensive-to-get, more objective binary masks. In our annotation experiment, we show that labeling light reflections is far more challenging and complex than conventional objects. Also, we show that our method can improve inter-annotator agreement and filter out annotator subjectivity. We train several State-Of-The-Art (SOTA) neural networks for semantic segmentation to demonstrate that our annotations can be used to detect light reflections from oncoming vehicles in urban scenarios before they are directly visible.

Machine learning-based systems are now part of a wide array of real-world applications seamlessly embedded in the social realm.In the wake of this realisation, strict legal regulations for these systems are currently being developed, addressing some of the risks they may pose.This is the coming of age of the interpretability and explainability problems in machine learning-based data analysis, which can no longer be seen just as an academic research problem.In this tutorial, associated to ESANN 2021 special session on "Interpretable Models in Machine Learning and Explainable Artificial Intelligence", we discuss explainable and interpretable machine learning as post-hoc and ante-hoc strategies to address these problems and highlight several aspects related to them, including their assessment.The contributions accepted for the session are then presented in this context.* A.V. is supported by Spanish

Deep learning struggles to generalize well to an unseen target domain of interest.Current domain adaptation methods simultaneously learn a classifier and an adversarial game for invariant representations but inadequately align local structures, while the underlying process is hard to interpret.We propose a new interpretable adversarial domain architecture, matching local manifold approximations across domains.Evaluated against related networks, the approach is competitive, while the adaptation process can be visually verified.

In current object detection, algorithms require the object to be directly visible in order to be detected. As humans, however, we intuitively use visual cues caused by the respective object to already make assumptions about its appearance. In the context of driving, such cues can be shadows during the day and often light reflections at night. In this paper, we study the problem of how to map this intuitive human behavior to computer vision algorithms to detect oncoming vehicles at night just from the light reflections they cause by their headlights. For that, we present an extensive open-source dataset containing 59 746 annotated grayscale images out of 346 different scenes in a rural environment at night. In these images, all oncoming vehicles, their corresponding light objects (e. g., headlamps), and their respective light reflections (e. g., light reflections on guardrails) are labeled. In this context, we discuss the characteristics of the dataset and the challenges in objectively describing visual cues such as light reflections. We provide different metrics for different ways to approach the task and report the results we achieved using state-of-the-art and custom object detection models as a first benchmark. With that, we want to bring attention to a new and so far neglected field in computer vision research, encourage more researchers to tackle the problem, and thereby further close the gap between human performance and computer vision systems.

Research on localization and perception for Autonomous Driving is mainly focused on camera and LiDAR datasets, rarely on radar data. Manually labeling sparse radar point clouds is challenging. For a dataset generation, we propose the cross sensor Radar Artifact Labeling Framework (RALF). Automatically generated labels for automotive radar data help to cure radar shortcomings like artifacts for the application of artificial intelligence. RALF provides plausibility labels for radar raw detections, distinguishing between artifacts and targets. The optical evaluation backbone consists of a generalized monocular depth image estimation of surround view cameras plus LiDAR scans. Modern car sensor sets of cameras and LiDAR allow to calibrate image-based relative depth information in overlapping sensing areas. K-Nearest Neighbors matching relates the optical perception point cloud with raw radar detections. In parallel, a temporal tracking evaluation part considers the radar detections' transient behavior. Based on the distance between matches, respecting both sensor and model uncertainties, we propose a plausibility rating of every radar detection. We validate the results by evaluating error metrics on semi-manually labeled ground truth dataset of $3.28\cdot10^6$ points. Besides generating plausible radar detections, the framework enables further labeled low-level radar signal datasets for applications of perception and Autonomous Driving learning tasks.

Visual perception is one of the most important information sources during driving. However, current camera perception systems are limited to object detection and, hence, to directly visible objects. Because it is mandatory for vehicles to run headlights at night, their emitted light can be detected before a vehicle is directly visible. Humans use this phenomenon to providently detect vehicles at night. In this paper, we analyze the discrepancy between ordinary vehicle detection and provident detection by quantifying the time gap between the two. This is achieved by conducting a test group study where participants are recorded while driving at night. Additionally, the dataset recorded during the study is used to provide a training and test dataset for machine learning approaches. To make use of the dataset for training machine learning methods, we analyze and discuss several annotation techniques. In a proof-of-concept, we used this dataset with its annotations to train a neural network on the direct and provident vehicle detection task. The resulting model shows that neural networks can successfully learn how to detect light-features. With further research and improvements, we are confident that a model for provident vehicle detection can be industrialized for use in production vehicles so that this information can be used in various safety and planning functions, including automatically adapting the high beam before it blinds other road users.

Methods for adversarial robustness certification aim to provide an upper bound on the test error of a classifier under adversarial manipulation of its input. Current certification methods are computationally expensive and limited to attacks that optimize the manipulation with respect to a norm. We overcome these limitations by investigating the robustness properties of Nearest Prototype Classifiers (NPCs) like learning vector quantization and large margin nearest neighbor. For this purpose, we study the hypothesis margin. We prove that if NPCs use a dissimilarity measure induced by a seminorm, the hypothesis margin is a tight lower bound on the size of adversarial attacks and can be calculated in constant time—this provides the first adversarial robustness certificate calculable in reasonable time. Finally, we show that each NPC trained by a triplet loss maximizes the hypothesis margin and is therefore optimized for adversarial robustness. In the presented evaluation, we demonstrate that NPCs optimized for adversarial robustness are competitive with state-of-the-art methods and set a new benchmark with respect to computational complexity for robustness certification.

Abstract Neural networks are state-of-the-art classification approaches but are generally difficult to interpret. This issue can be partly alleviated by constructing a precise decision process within the neural network. In this work, a network architecture, denoted as Classification-By-Components network (CBC), is proposed. It is restricted to follow an intuitive reasoning based decision process inspired by Biederman's recognition-by-components theory from cognitive psychology. The network is trained to learn and detect generic components that characterize objects. In parallel, a class-wise reasoning strategy based on these components is learned to solve the classification problem. In contrast to other work on reasoning, we propose three different types of reasoning: positive, negative, and indefinite. These three types together form a probability space to provide a probabilistic classifier. The decomposition of objects into generic components combined with the probabilistic reasoning provides by design a clear interpretation of the classification decision process. The evaluation of the approach on MNIST shows that CBCs are viable classifiers. Additionally, we demonstrate that the inherent interpretability offers a profound understanding of the classification behavior such that we can explain the success of an adversarial attack. The method's scalability is successfully tested using the ImageNet dataset.

The paper deals with realizations of transfer learning for classification, i. e. the adaptation of a classifier model to a changed data distribution. This change could be a data drift or a more complex transformation. We propose to model those data changes by manifolds describing continuous transformations of the data. This description can be seen as a generalization of function based transfer models considered so far. The manifold description of the transfer function allows either to adjust the classifier model to the changed data distribution or a back-transformation of those data to the original data space. To get the approach feasible, the manifold is approximated by the affine part of the Taylor expansion of the manifold structure. Moreover, the affine approximation shows mathematical correspondences to tangent metric leaning, which was developed for handling of data with drifts in classification methods. The paper provides the mathematical background for manifold based transfer data learning. Further, the approach is exemplarily applied for the generalized learning vector quantization classifier. This classifier is a prominent method which frequently achieves a high performance and a robust behavior while the classifier complexity is low compared to more sophisticated approaches like deep learning architectures or support vector machines. Moreover, the good interpretability of learning vector quantization classifiers also contributes to an intuitive practical understanding of transfer learning.

We propose a learning vector quantization algorithm variant for prototype-based classification learning with adaptive tangent distance learning. Tangent distances were developed to achieve dissimilarity measures invariant with respect to transformations and distortions like rotation, noise, etc.. Usually, these tangent distances are predefined in applications or are estimated in preprocessing. We introduce in this paper a generalized learning vector quantizer (GLVQ) with an online adaptation scheme for tangent distances. The adaptation takes place as a stochastic gradient descent learning accompanying the usual online prototype learning. In this way, class discriminative tangents are learned contributing to a better classification performance. Further, the resulting update schemes can be seen as a special type of local matrix learning in GLVQ. In this paper, we provide the full mathematical theory behind the derived tangent distance adaptation rule and demonstrate the classification ability of the resulting GLVQ model in comparison to state-of-the-art tangent distance based classifiers in the field.