Convolutional neural networks have been used to achieve a string of successes during recent years, but their lack of interpretability remains a serious issue. Adversarial examples are designed to deliberately fool neural networks into making any desired incorrect classification, potentially with very high certainty. Several defensive approaches increase robustness against adversarial attacks, demanding attacks of greater magnitude, which lead to visible artifacts. By considering human visual perception, we compose a technique that allows to hide such adversarial attacks in regions of high complexity, such that they are imperceptible even to an astute observer. We carry out a user study on classifying adversarially modified images to validate the perceptual quality of our approach and find significant evidence for its concealment with regards to human visual perception.

Abstract Convolutional neural networks (CNNs) are deep learning frameworks which are well known for their notable performance in classification tasks. Hence, many skeleton-based action recognition and segmentation (SBARS) algorithms benefit from them in their designs. However, a shortcoming of such applications is the general lack of spatial relationships between the input features in such data types. Besides, non-uniform temporal scalings are a common issue in skeleton-based data streams which leads to having different input sizes even within one specific action category. In this work, we propose a novel deep-aligned convolutional neural network (DACNN) to tackle the above challenges for the particular problem of SBARS. Our network is designed by introducing a new type of filters in the context of CNNs which are trained based on their alignments to the local subsequences in the inputs. These filters result in efficient predictions as well as learning interpretable patterns in the data. Also, our DACNN framework can incrementally expand its deep structure based on the learning progress, which makes it flexible regarding different SBARS datasets. We empirically evaluate our framework on real-world benchmarks showing that the proposed DACNN algorithm obtains a competitive performance compared to the state of the art while benefiting from a less complicated yet more interpretable model.

Prototype-based methods are of the particular interest for domain specialists and practitioners as they summarize a dataset by a small set of representatives. Therefore, in a classification setting, interpretability of the prototypes is as significant as the prediction accuracy of the algorithm. Nevertheless, the state-of-the-art methods make inefficient trade-offs between these concerns by sacrificing one in favor of the other, especially if the given data has a kernel-based (or multiple-kernel) representation. In this paper, we propose a novel interpretable multiple-kernel prototype learning (IMKPL) to construct highly interpretable prototypes in the feature space, which are also efficient for the discriminative representation of the data. Our method focuses on the local discrimination of the classes in the feature space and shaping the prototypes based on condensed class-homogeneous neighborhoods of data. Besides, IMKPL learns a combined embedding in the feature space in which the above objectives are better fulfilled. When the base kernels coincide with the data dimensions, this embedding results in a discriminative features selection. We evaluate IMKPL on several benchmarks from different domains which demonstrate its superiority to the related state-of-the-art methods regarding both interpretability and discriminative representation.

In recent years, kernel-based sparse coding (K-SRC) has received a special attention due to its efficient representation of nonlinear data structures in the feature space. Nevertheless, the existing K-SRC methods suffer from the lack of consistency between their training and test optimization frameworks. In this work, we propose a novel confident K-SRC and dictionary learning algorithm (CKSC) which focuses on the discriminative reconstruction of the data based on its representation in the kernel space. CKSC focuses on reconstructing each data sample via weighted contributions which are confident in its corresponding class of data. We employ novel discriminative terms to apply this scheme to both training and test frameworks in our algorithm. This increases the consistency of these optimization frameworks and improves the discriminative performance in the recall phase. In addition, CKSC directly employs the supervised information in its dictionary learning framework to enhance the discriminative structure of the dictionary. For empirical evaluations, we implement our CKSC algorithm on multivariate time-series benchmarks such as DynTex++ and UTKinect. Our claims regarding the superior performance of the proposed algorithm are justified throughout comparing its classification results to the state-of-the-art K-SRC algorithms.

An increasing number of industrial areas recognize the opportunities of Big Data, requiring highly efficient algorithms which enable real-time processing to reduce the burden of data storage and maintenance. Decision trees are extremely fast, highly accurate and easy to use in practice. Merging multiple decision trees to an ensemble leads to one of the most powerful machine learning methods. The Very Fast Decision Tree is the state-of-the-art incremental decision tree induction algorithm, capable of learning from massive data streams. It is successful due to its theoretical guarantees based on the Hoeffding bound as well as its competitive performance in terms of classification accuracy and time / space efficiency. In this paper, we increase the efficiency even further by replacing its global splitting scheme, which periodically tries to split every nminexamples. Instead, we utilize local statistics to predict the split-time, thus, avoiding unnecessary split-attempts, usually dominating the computational cost. Concretely, we use the class distributions of previous split-attempts to approximate the minimum number of examples until the Hoeffding bound is met. This cautious approach yields by design a low delay and reduces the number of split-attempts at the same time. We extensively evaluate our method using common stream-learning benchmarks also considering non-stationary environments. The experiments confirm a substantially reduced run-time without a loss in classification performance.

Data Mining in non-stationary data streams is gaining more attentionrecently, especially in the context of Internet of Things and Big Data. It is a highly challenging task, since the fundamentally different typesof possibly occurring drift undermine classical assumptions such asi.i.d. data or stationary distributions. Available algorithms are either struggling with certain forms of drift or require a priori knowledge in terms of a task specific setting. We propose the Self Adjusting Memory (SAM) model for the k Nearest Neighbor (kNN) algorithm since kNN constitutes a proven classifier within the streaming setting. SAM-kNN can deal with heterogeneous concept drift, i.e different drift types and rates, using biologically inspiredmemory models and their coordination. It can be easilyapplied in practice since an optimization of the meta parameters is not necessary. The basic idea is to construct dedicated models for thecurrent and former concepts and apply them according tothe demands of the given situation. An extensive evaluation on various benchmarks, consisting of artificial streamswith known drift characteristics as well as real world datasets is conducted. Thereby, we explicitly add new benchmarks enabling a precise performance evaluation on multiple types of drift. The highly competitive results throughout all experiments underline the robustness of SAM-kNN as well as its capabilityto handle heterogeneous concept drift.

We investigate metric learning in the context of dynamic time warping (DTW), the by far most popular dissimilarity measure used for the comparison and analysis of motion capture data. While metric learning enables a problem-adapted representation of data, the majority of methods has been proposed for vectorial data only. In this contribution, we extend the popular principle offered by the large margin nearest neighbours learner (LMNN) to DTW by treating the resulting component-wise dissimilarity values as features. We demonstrate, that this principle greatly enhances the classification accuracy in several benchmarks. Further, we show that recent auxiliary concepts such as metric regularisation can be transferred from the vectorial case to component-wise DTW in a similar way. We illustrate, that metric regularisation constitutes a crucial prerequisite for the interpretation of the resulting relevance profiles.

Linear data transformations constitute essential operations in various machine learning algorithms, ranging from linear regression up to adaptive metric transformation. Often, linear scalings are not only used to improve the model accuracy, rather feature coefficients as provided by the mapping are interpreted as an indicator for the relevance of the feature for the task at hand. This principle, however, can be misleading in particular for high-dimensional or correlated features, since it easily marks irrelevant features as relevant or vice versa. In this contribution, we propose a mathematical formalisation of the minimum and maximum feature relevance for a given linear transformation which can efficiently be solved by means of linear programming. We evaluate the method in several benchmarks, where it becomes apparent that the minimum and maximum relevance closely resembles what is often referred to as weak and strong relevance of the features; hence unlike the mere scaling provided by the linear mapping, it ensures valid interpretability.

Linear data transformations are essential operations in many machine learning algorithms, helping to make such models more flexible or to emphasize certain data directions. In particular for high dimensional data sets linear transformations are not necessarily uniquely determined, though, and alternative parameterizations exist which do not change the mapping of the training data. Thus, regularization is required to make the model robust to noise and more interpretable for the user. In this contribution, we characterize the group of transformations which leave a linear mapping invariant for a given finite data set, and we discuss the consequences on the interpretability of the models. We propose an intuitive regularization mechanism to avoid problems in under-determined configurations, and we test the approach in two machine learning models.

A wealth of powerful dimensionality reduction methods has been established which can be used for data visualization and preprocessing. These are accompanied by formal evaluation schemes, which allow a quantitative evaluation along general principles and which even lead to further visualization schemes based on these objectives. Most methods, however, provide a mapping of a priorly given finite set of points only, requiring additional steps for out-of-sample extensions. We propose a general view on dimensionality reduction based on the concept of cost functions, and, based on this general principle, extend dimensionality reduction to explicit mappings of the data manifold. This offers simple out-of-sample extensions. Further, it opens a way towards a theory of data visualization taking the perspective of its generalization ability to new data points. We demonstrate the approach based on a simple global linear mapping as well as prototype-based local linear mappings.

The information theoretical learnability of folding networks, a very successful approach capable of dealing with tree structured inputs, is examined. We find bounds on the VC, pseudo-, and fat shattering dimension of folding networks with various activation functions. As a consequence, valid generalization of folding networks can be guaranteed. However, distribution independent bounds on the generalization error cannot exist in principle. We propose two approaches which take the specific distribution into account and allow us to derive explicit bounds on the deviation of the empirical error from the real error of a learning algorithm. The first approach requires the probability of large trees to be limited a priori and the second approach deals with situations where the maximum input height in a concrete learning example is restricted.