Search

We consider the problem of learning with sensitive features under the privileged information setting where the goal is to learn a classifier that uses features not available (or too sensitive to collect) at test/deployment time to learn a better model at training time. We focus on tree-based learners, specifically gradient-boosted decision trees for learning with privileged information. Our methods use privileged features as knowledge to guide the algorithm when learning from fully observed (usable) features. We derive the theory, empirically validate the effectiveness of our algorithms, and verify them on standard fairness metrics.

The number of satellites in orbit around Earth is increasing rapidly, with the risk of collision rising accordingly. Trends of the global population of satellites need to be analyzed to test the viability and impact of proposed rules and laws affecting the satellite population and collision avoidance strategies. This requires large scale simulations of satellites that are propagated on long timescales to compute the large amounts of actionable close encounters (called conjunctions), which could lead to collisions. Rigorously checking for conjunctions by computing future states of orbits is computationally expensive due to the large amount of objects involved and conjunction filters are thus used to remove non-conjuncting orbit pairs from the list of possible conjunctions. In this work, we explore the possibility of machine learning (ML) based conjunction filters using several algorithms such as eXtreme Gradient Boosting, TabNet and (physics-informed) neural networks and deep operator networks. To show the viability and the potential of ML based filters, these algorithms are trained to predict the future state of orbits. For the physics-informed approaches, multiple partial differential equations are set up using the Kepler equation as a basis. The empirical results demonstrate that physics-informed deep operator networks are capable of predicting the future state of orbits using these equations (RMSE: 0.136) and outperform eXtreme Gradient Boosting (RMSE: 0.568) and TabNet (RMSE: 0.459). We also propose a filter based on the trained deep operator network which is shown to outperforms the filter capability of the commonly used perigee-apogee test and the orbit path filter on a synthetic dataset, while being on average 3.2 times faster to compute than a rigorous conjunction check.

Natural gas pipelines represent a critical infrastructure for most countries and thus their safety is of paramount importance. To report potential risks along pipelines, several steps are taken such as manual inspection and helicopter flights; however, these solutions are expensive and the flights are environmentally unfriendly. Deep learning has demonstrated considerable potential in handling a number of tasks in recent years as models rely on huge datasets to learn a specific task. With the increasing number of satellites orbiting the Earth, remote sensing data have become widely available, thus paving the way for automated pipeline monitoring via deep learning. This can result in effective risk detection, thereby reducing monitoring costs while being more precise and accurate. A major hindrance here is the low resolution of images obtained from the satellites, which makes it difficult to detect smaller changes. To this end, we propose to use transformers trained with low-resolution images in a change detection setting to detect pipeline risks. We collect PlanetScope satellite imagery (3 m resolution) that captures certain risks associated with the pipelines and present how we collected the data. Furthermore, we compare various state-of-the-art models, among which ChangeFormer, a transformer architecture for change detection, achieves the best performance with a 70% F1 score. As part of our evaluation, we discuss the specific performance requirements in pipeline monitoring and show how the model’s predictions can be shifted accordingly during training.

In recent years, deep neural networks for strategy games have made significant progress. AlphaZero-like frameworks which combine Monte-Carlo tree search with reinforcement learning have been successfully applied to numerous games with perfect information. However, they have not been developed for domains where uncertainty and unknowns abound, and are therefore often considered unsuitable due to imperfect observations. Here, we challenge this view and argue that they are a viable alternative for games with imperfect information—a domain currently dominated by heuristic approaches or methods explicitly designed for hidden information, such as oracle-based techniques. To this end, we introduce a novel algorithm based solely on reinforcement learning, called AlphaZe∗∗, which is an AlphaZero-based framework for games with imperfect information. We examine its learning convergence on the games Stratego and DarkHex and show that it is a surprisingly strong baseline, while using a model-based approach: it achieves similar win rates against other Stratego bots like Pipeline Policy Space Response Oracle (P2SRO), while not winning in direct comparison against P2SRO or reaching the much stronger numbers of DeepNash. Compared to heuristics and oracle-based approaches, AlphaZe∗∗ can easily deal with rule changes, e.g., when more information than usual is given, and drastically outperforms other approaches in this respect.

As wind is the basis of all wind energy projects, a precise knowledge about its availability is needed. For an analysis of the site-specific wind conditions, Virtual Meteorological Masts (VMMs) are frequently used. VMMs make use of site calibrated numerical data to provide precise wind estimates during all phases of a wind energy project. Typically, numerical data are used for the long-term correlation that is required for estimating the yield of new wind farm projects. However, VMMs can also be used to fill data gaps or during the operational phase as an additional reference data set to detect degrading sensors. The value of a VMM directly depends on its ability and precision to reproduce site-specific environmental conditions. Commonly, linear regression is used as state of the art to correct reference data to the site-specific conditions. In this study, a framework of 10 different machine-learning methods is tested to investigated the benefit of more advanced methods on two offshore and one onshore site. We find significantly improving correlations between the VMMs and the reference data when using more advanced methods and present the most promising ones. The K-Nearest Neighbors and AdaBoost regressors show the best results in our study, but Multi-Output Mixture of Gaussian Processes is also very promising. The use of more advanced regression models lead to decreased uncertainties; hence those methods should find its way into industrial applications. The recommended regression models can serve as a starting point for the development of end-user applications and services.

Data-driven approaches are becoming increasingly common as problem-solving tools in many areas of science and technology. In most cases, machine learning models are the key component of these solutions. Often, a solution involves multiple learning models, along with significant levels of reasoning with the models' output and input. However, the current tools are cumbersome not only for domain experts who are not fluent in machine learning but also for machine learning experts who evaluate new algorithms and models on real-world data and develop AI systems. We review key efforts made by various AI communities in providing languages for high-level abstractions over learning and reasoning techniques needed for designing complex AI systems. We classify the existing frameworks based on the type of techniques and their data and knowledge representations, compare the ways the current tools address the challenges of programming real-world applications and highlight some shortcomings and future directions. Our comparison is only qualitative and not experimental since the performance of the systems is not a factor in our study.

The ability of a robot to generate appropriate facial expressions is a key aspect of perceived sociability in human-robot interaction. Yet many existing approaches rely on the use of a set of fixed, preprogrammed joint configurations for expression generation. Automating this process provides potential advantages to scale better to different robot types and various expressions. To this end, we introduce ExGenNet, a novel deep generative approach for facial expressions on humanoid robots. ExGenNets connect a generator network to reconstruct simplified facial images from robot joint configurations with a classifier network for state-of-the-art facial expression recognition. The robots’ joint configurations are optimized for various expressions by backpropagating the loss between the predicted expression and intended expression through the classification network and the generator network. To improve the transfer between human training images and images of different robots, we propose to use extracted features in the classifier as well as in the generator network. Unlike most studies on facial expression generation, ExGenNets can produce multiple configurations for each facial expression and be transferred between robots. Experimental evaluations on two robots with highly human-like faces, Alfie (Furhat Robot) and the android robot Elenoide, show that ExGenNet can successfully generate sets of joint configurations for predefined facial expressions on both robots. This ability of ExGenNet to generate realistic facial expressions was further validated in a pilot study where the majority of human subjects could accurately recognize most of the generated facial expressions on both the robots.

Classification approaches that allow to extract logical rules such as decision trees are often considered to be more interpretable than neural networks. Also, logical rules are comparatively easy to verify with any possible input. This is an important part in systems that aim to ensure correct operation of a given model. However, for high-dimensional input data such as images, the individual symbols, i.e. pixels, are not easily interpretable. Therefore, rule-based approaches are not typically used for this kind of high-dimensional data. We introduce the concept of first-order convolutional rules, which are logical rules that can be extracted using a convolutional neural network (CNN), and whose complexity depends on the size of the convolutional filter and not on the dimensionality of the input. Our approach is based on rule extraction from binary neural networks with stochastic local search. We show how to extract rules that are not necessarily short, but characteristic of the input, and easy to visualize. Our experiments show that the proposed approach is able to model the functionality of the neural network while at the same time producing interpretable logical rules. Thus, we demonstrate the potential of rule-based approaches for images which allows to combine advantages of neural networks and rule learning.

Allowing machines to choose whether to kill humans would be devastating for world peace and security. But how do we equip machines with the ability to learn ethical or even moral choices? In this study, we show that applying machine learning to human texts can extract deontological ethical reasoning about “right” and “wrong” conduct. We create a template list of prompts and responses, such as “Should I [action]?”, “Is it okay to [action]?”, etc. with corresponding answers of “Yes/no, I should (not).” and "Yes/no, it is (not)." The model's bias score is the difference between the model's score of the positive response (“Yes, I should”) and that of the negative response (“No, I should not”). For a given choice, the model's overall bias score is the mean of the bias scores of all question/answer templates paired with that choice. Specifically, the resulting model, called the Moral Choice Machine (MCM), calculates the bias score on a sentence level using embeddings of the Universal Sentence Encoder since the moral value of an action to be taken depends on its context. It is objectionable to kill living beings, but it is fine to kill time. It is essential to eat, yet one might not eat dirt. It is important to spread information, yet one should not spread misinformation. Our results indicate that text corpora contain recoverable and accurate imprints of our social, ethical and moral choices, even with context information. Actually, training the Moral Choice Machine on different temporal news and book corpora from the year 1510 to 2008/2009 demonstrate the evolution of moral and ethical choices over different time periods for both atomic actions and actions with context information. By training it on different cultural sources such as the Bible and the constitution of different countries, the dynamics of moral choices in culture, including technology are revealed. That is the fact that moral biases can be extracted, quantified, tracked, and compared across cultures and over time.

Deep neural networks have been successfully applied in learning the board games Go, chess, and shogi without prior knowledge by making use of reinforcement learning. Although starting from zero knowledge has been shown to yield impressive results, it is associated with high computationally costs especially for complex games. With this paper, we present CrazyAra which is a neural network based engine solely trained in supervised manner for the chess variant crazyhouse. Crazyhouse is a game with a higher branching factor than chess and there is only limited data of lower quality available compared to AlphaGo. Therefore, we focus on improving efficiency in multiple aspects while relying on low computational resources. These improvements include modifications in the neural network design and training configuration, the introduction of a data normalization step and a more sample efficient Monte-Carlo tree search which has a lower chance to blunder. After training on 569537 human games for 1.5 days we achieve a move prediction accuracy of 60.4%. During development, versions of CrazyAra played professional human players. Most notably, CrazyAra achieved a four to one win over 2017 crazyhouse world champion Justin Tan (aka LM Jann Lee) who is more than 400 Elo higher rated compared to the average player in our training set. Furthermore, we test the playing strength of CrazyAra on CPU against all participants of the second Crazyhouse Computer Championships 2017, winning against twelve of the thirteen participants. Finally, for CrazyAraFish we continue training our model on generated engine games. In 10 long-time control matches playing Stockfish 10, CrazyAraFish wins three games and draws one out of 10 matches.

Deep learning is bringing remarkable contributions to the field of argumentation mining, but the existing approaches still need to fill the gap toward performing advanced reasoning tasks. In this position paper, we posit that neural-symbolic and statistical relational learning could play a crucial role in the integration of symbolic and sub-symbolic methods to achieve this goal.

Wound healing is a complex and dynamic process with different distinct and overlapping phases from homeostasis, inflammation and proliferation to remodelling. Monitoring the healing response of injured tissue is of high importance for basic research and clinical practice. In traditional application, biological markers characterize normal and abnormal wound healing. Understanding functional relationships of these biological processes is essential for developing new treatment strategies. However, most of the present techniques (in vitro or in vivo) include invasive microscopic or analytical tissue sampling. In the present study, a non-invasive alternative for monitoring processes during wound healing is introduced. Within this context, hyperspectral imaging (HSI) is an emerging and innovative non-invasive imaging technique with different opportunities in medical applications. HSI acquires the spectral reflectance of an object, depending on its biochemical and structural characteristics. An in-vitro 3-dimensional (3-D) wound model was established and incubated without and with acute and chronic wound fluid (AWF, CWF), respectively. Hyperspectral images of each individual specimen of this 3-D wound model were assessed at day 0/5/10 in vitro, and reflectance spectra were evaluated. For analysing the complex hyperspectral data, an efficient unsupervised approach for clustering massive hyperspectral data was designed, based on efficient hierarchical decomposition of spectral information according to archetypal data points. It represents, to the best of our knowledge, the first application of an advanced Data Mining approach in context of non-invasive analysis of wounds using hyperspectral imagery. By this, temporal and spatial pattern of hyperspectral clusters were determined within the tissue discs and among the different treatments. Results from non-invasive imaging were compared to the number of cells in the various clusters, assessed by Hematoxylin/Eosin (H/E) staining. It was possible to correlate cell quantity and spectral reflectance during wound closure in a 3-D wound model in vitro.

Previous plant phenotyping studies have focused on the visible (VIS, 400−700 nm), near-infrared (NIR, 700−1000 nm) and short-wave infrared (SWIR, 1000−2500 nm) range. The ultraviolet range (UV, 200−380 nm) has not yet been used in plant phenotyping even though a number of plant molecules like flavones and phenol feature absorption maxima in this range. In this study an imaging UV line scanner in the range of 250−430 nm is introduced to investigate crop plants for plant phenotyping. Observing plants in the UV-range can provide information about important changes of plant substances. To record reliable and reproducible time series results, measurement conditions were defined that exclude phototoxic effects of UV-illumination in the plant tissue. The measurement quality of the UV-camera has been assessed by comparing it to a non-imaging UV-spectrometer by measuring six different plant-based substances. Given the findings of these preliminary studies, an experiment has been defined and performed monitoring the stress response of barley leaves to salt stress. The aim was to visualize the effects of abiotic stress within the UV-range to provide new insights into the stress response of plants. Our study demonstrated the first use of a hyperspectral sensor in the UV-range for stress detection in plant phenotyping.

Understanding the response dynamics of plants to biotic stress is essential to improve management practices and breeding strategies of crops and thus to proceed towards a more sustainable agriculture in the coming decades. In this context, hyperspectral imaging offers a particularly promising approach since it provides non-destructive measurements of plants correlated with internal structure and biochemical compounds. In this paper, we present a cascade of data mining techniques for fast and reliable data-driven sketching of complex hyperspectral dynamics in plant science and plant phenotyping. To achieve this, we build on top of a recent linear time matrix factorization technique, called Simplex Volume Maximization, in order to automatically discover archetypal hyperspectral signatures that are characteristic for particular diseases. The methods were applied on a data set of barley leaves (Hordeum vulgare) diseased with foliar plant pathogens Pyrenophora teres, Puccinia hordei and Blumeria graminis hordei. Towards more intuitive visualizations of plant disease dynamics, we use the archetypal signatures to create structured summaries that are inspired by metro maps, i.e. schematic diagrams of public transport networks. Metro maps of plant disease dynamics produced on several real-world data sets conform to plant physiological knowledge and explicitly illustrate the interaction between diseases and plants. Most importantly, they provide an abstract and interpretable view on plant disease progression.