Hessian Center for Artificial Intelligence

Apple recently revealed its deep perceptual hashing system NeuralHash to detect child sexual abuse material (CSAM) on user devices before files are uploaded to its iCloud service. Public criticism quickly arose regarding the protection of user privacy and the system’s reliability. In this paper, we present the first comprehensive empirical analysis of deep perceptual hashing based on NeuralHash. Specifically, we show that current deep perceptual hashing may not be robust. An adversary can manipulate the hash values by applying slight changes in images, either induced by gradient-based approaches or simply by performing standard image transformations, forcing or preventing hash collisions. Such attacks permit malicious actors easily to exploit the detection system: from hiding abusive material to framing innocent users, everything is possible. Moreover, using the hash values, inferences can still be made about the data stored on user devices. In our view, based on our results, deep perceptual hashing in its current form is generally not ready for robust client-side scanning and should not be used from a privacy perspective.1

Generative AI models have recently achieved astonishing results in quality and are consequently employed in a fast-growing number of applications. However, since they are highly data-driven, relying on billion-sized datasets randomly scraped from the internet, they also suffer from degenerated and biased human behavior, as we demonstrate. In fact, they may even reinforce such biases. To not only uncover but also combat these undesired effects, we present a novel strategy, called Fair Diffusion, to attenuate biases during the deployment of generative text-to-image models. Specifically, we demonstrate shifting a bias in any direction based on human instructions yielding arbitrary proportions for, e.g., identity groups. As our empirical evaluation demonstrates, this introduced control enables instructing generative image models on fairness, requiring no data filtering nor additional training.

In this paper, we introduce an innovative approach to handling the multi-armed bandit (MAB) problem in non-stationary environments, harnessing the predictive power of large language models (LLMs). With the realization that traditional bandit strategies, including epsilon-greedy and upper confidence bound (UCB), may struggle in the face of dynamic changes, we propose a strategy informed by LLMs that offers dynamic guidance on exploration versus exploitation, contingent on the current state of the bandits. We bring forward a new non-stationary bandit model with fluctuating reward distributions and illustrate how LLMs can be employed to guide the choice of bandit amid this variability. Experimental outcomes illustrate the potential of our LLM-informed strategy, demonstrating its adaptability to the fluctuating nature of the bandit problem, while maintaining competitive performance against conventional strategies. This study provides key insights into the capabilities of LLMs in enhancing decision-making processes in dynamic and uncertain scenarios.

Katharina Haemmerl, Bjoern Deiseroth, Patrick Schramowski, Jindřich Libovický, Constantin Rothkopf, Alexander Fraser, Kristian Kersting. Findings of the Association for Computational Linguistics: ACL 2023. 2023.

We review the state of research on empathy in natural language processing and identify the following issues: (1) empathy definitions are absent or abstract, which (2) leads to low construct validity and reproducibility. Moreover, (3) emotional empathy is overemphasized, skewing our focus to a narrow subset of simplified tasks. We believe these issues hinder research progress and argue that current directions will benefit from a clear conceptualization that includes operationalizing cognitive empathy components. Our main objectives are to provide insight and guidance on empathy conceptualization for NLP research objectives and to encourage researchers to pursue the overlooked opportunities in this area, highly relevant, e.g., for clinical and educational sectors.

Active exploration of the visual environment requires choosing the next gaze target, which has been characterized as depending on attentional biases and on oculomotor biases. However, disentangling these two factors has provided contradictory results. Here, we conceptualize active gaze selection as a decision-making process in the context of neuroeconomics, allowing us to employ a new saccadic preference elicitation experiment and to quantify behavioral biases with a random utility model. Subjects showed similar preferences for shorter amplitude saccades in line with previous research and shared a preference for saccades back towards the previous gaze location, contrary to numerous reports on spatial “inhibition of return”. Surprisingly, the experiment revealed marked individual differences in the preferences for global gaze directions, which were stable across most participants after one week. Thus, this study provides the first measurement of intrinsic saccadic utilities independently of spatial image properties and quantitatively dissociates stable individual differences in global gaze selection.

To navigate through their immediate environment humans process scene information rapidly. How does the cascade of neural processing elicited by scene viewing to facilitate navigational planning unfold over time? To investigate, we recorded human brain responses to visual scenes with electroencephalography and related those to computational models that operationalize three aspects of scene processing (2D, 3D, and semantic information), as well as to a behavioral model capturing navigational affordances. We found a temporal processing hierarchy: navigational affordance is processed later than the other scene features (2D, 3D, and semantic) investigated. This reveals the temporal order with which the human brain computes complex scene information and suggests that the brain leverages these pieces of information to plan navigation.

Normative and descriptive models have long vied to explain and predict human risky choices, such as those between goods or gambles. A recent study reported the discovery of a new, more accurate model of human decision-making by training neural networks on a new online large-scale dataset, choices13k. Here we systematically analyse the relationships between several models and datasets using machine-learning methods and find evidence for dataset bias. Because participants' choices in stochastically dominated gambles were consistently skewed towards equipreference in the choices13k dataset, we hypothesized that this reflected increased decision noise. Indeed, a probabilistic generative model adding structured decision noise to a neural network trained on data from a laboratory study transferred best, that is, outperformed all models apart from those trained on choices13k. We conclude that a careful combination of theory and data analysis is still required to understand the complex interactions of machine-learning models and data of human risky choices.

Anomaly detection is essential for realising state-of-the-art and safe cyber-physical production systems. With the early detection of possible anomalies, the rise and the carryover of a failure throughout the entire manufacture can be prevented. Current anomaly detection approaches use sensor data from the live system only. Therefore, many important characteristics, interactions, and correlations within the system are neglected. Graphs open a particular way of modelling cyber-physical production systems, while at the same time, they allow the inclusion of a wider variation and variety of data about the system. In this study, we introduce an unsupervised correlation- and interaction-aware anomaly detection concept for cyber-physical production systems based on graph neural networks. We define a concept that allows modelling the entire system as a graph, including external knowledge. We introduce a reconstruction-based graph neural network to detect and analyse the anomalies in these graph structures. The concept is evaluated in a real industrial cyber-physical production system. The test results confirm that the presented concept can be applied to detect anomalies while concurrently considering the correlations and interactions of the system.

Shared dynamics models are important for capturing the complexity and variability inherent in Human-Robot Interaction (HRI). Therefore, learning such shared dynamics models can enhance coordination and adaptability to enable successful reactive interactions with a human partner. In this work, we propose a novel approach for learning a shared latent space representation for HRIs from demonstrations in a Mixture of Experts fashion for reactively generating robot actions from human observations. We train a Variational Autoencoder (VAE) to learn robot motions regularized using an informative latent space prior that captures the multimodality of the human observations via a Mixture Density Network (MDN). We show how our formulation derives from a Gaussian Mixture Regression formulation that is typically used approaches for learning HRI from demonstrations such as using an HMM/GMM for learning a joint distribution over the actions of the human and the robot. We further incorporate an additional regularization to prevent mode “collapse”, a common phenomenon when using latent space mixture models with VAEs. We find that our approach of using an informative MDN prior from human observations for a VAE generates more accurate robot motions compared to previous HMM-based or recurrent approaches of learning shared latent representations, which we validate on various HRI datasets involving interactions such as handshakes, fistbumps, waving, and handovers. Further experiments in a real-world human-to-robot handover scenario show the efficacy of our approach for generating successful interactions with four different human interaction partners.

EMOKINE is a software package and dataset creation suite for emotional full-body movement research in experimental psychology, affective neuroscience, and computer vision. A computational framework, comprehensive instructions, a pilot dataset, observer ratings, and kinematic feature extraction code are provided to facilitate future dataset creations at scale. In addition, the EMOKINE framework outlines how complex sequences of movements may advance emotion research. Traditionally, often emotional-'action'-based stimuli are used in such research, like hand-waving or walking motions. Here instead, a pilot dataset is provided with short dance choreographies, repeated several times by a dancer who expressed different emotional intentions at each repetition: anger, contentment, fear, joy, neutrality, and sadness. The dataset was simultaneously filmed professionally, and recorded using XSENS® motion capture technology (17 sensors, 240 frames/second). Thirty-two statistics from 12 kinematic features were extracted offline, for the first time in one single dataset: speed, acceleration, angular speed, angular acceleration, limb contraction, distance to center of mass, quantity of motion, dimensionless jerk (integral), head angle (with regards to vertical axis and to back), and space (convex hull 2D and 3D). Average, median absolute deviation (MAD), and maximum value were computed as applicable. The EMOKINE software is appliable to other motion-capture systems and is openly available on the Zenodo Repository. Releases on GitHub include: (i) the code to extract the 32 statistics, (ii) a rigging plugin for Python for MVNX file-conversion to Blender format (MVNX=output file XSENS® system), and (iii) a Python-script-powered custom software to assist with blurring faces; latter two under GPLv3 licenses.

The AlphaZero algorithm has been successfully applied in a range of discrete domains, most notably board games. It utilizes a neural network that learns a value and policy function to guide the exploration in a Monte-Carlo Tree Search. Although many search improvements such as graph search have been proposed for Monte-Carlo Tree Search in the past, most of them refer to an older variant of the Upper Confidence bounds for Trees algorithm that does not use a policy for planning. We improve the search algorithm for AlphaZero by generalizing the search tree to a directed acyclic graph. This enables information flow across different subtrees and greatly reduces memory consumption. Along with Monte-Carlo Graph Search, we propose a number of further extensions, such as the inclusion of Epsilon-Greedy exploration, a revised terminal solver and the integration of domain knowledge as constraints. In our empirical evaluations, we use the CrazyAra engine on chess and crazyhouse as examples to show that these changes bring significant improvements to AlphaZero.

In this work, we introduce composable energy policies (CEP), a novel framework for multi-objective motion generation. We frame the problem of composing multiple policy components from a probabilistic view. We consider a set of stochastic policies represented in arbitrary task spaces, where each policy represents a distribution of the actions to solve a particular task. Then, we aim to find the action in the configuration space that optimally satisfies all the policy components. The presented framework allows the fusion of motion generators from different sources: optimal control, data-driven policies, motion planning, and handcrafted policies. Classically, the problem of multi-objective motion generation is solved by the composition of a set of deterministic policies, rather than stochastic policies. However, there are common situations where different policy components have conflicting behaviors, leading to oscillations or the robot getting stuck in an undesirable state. While our approach is not directly able to solve the conflicting policies problem, we claim that modeling each policy as a stochastic policy allows more expressive representations for each component in contrast with the classical reactive motion generation approaches. In some tasks, such as reaching a target in a cluttered environment, we show experimentally that CEP additional expressivity allows us to model policies that reduce these conflicting behaviors. A field that benefits from these reactive motion generators is the one of robot reinforcement learning. Integrating these policy architectures with reinforcement learning allows us to include a set of inductive biases in the learning problem. These inductive biases guide the reinforcement learning agent towards informative regions or improve collision safety while exploring. In our work, we show how to integrate our proposed reactive motion generator as a structured policy for reinforcement learning. Combining the reinforcement learning agent exploration with the prior-based CEP, we can improve the learning performance and explore safer.

Local-first software manages and processes private data locally while still enabling collaboration between multiple parties connected via partially unreliable networks. Such software typically involves interactions with users and the execution environment (the outside world). The unpredictability of such interactions paired with their decentralized nature make reasoning about the correctness of local-first software a challenging endeavor. Yet, existing solutions to develop local-first software do not provide support for automated safety guarantees and instead expect developers to reason about concurrent interactions in an environment with unreliable network conditions. We propose LoRe , a programming model and compiler that automatically verifies developer-supplied safety properties for local-first applications. LoRe combines the declarative data flow of reactive programming with static analysis and verification techniques to precisely determine concurrent interactions that violate safety invariants and to selectively employ strong consistency through coordination where required. We propose a formalized proof principle and demonstrate how to automate the process in a prototype implementation that outputs verified executable code. Our evaluation shows that LoRe simplifies the development of safe local-first software when compared to state-of-the-art approaches and that verification times are acceptable.

It is expected that in the near future, AI software development assistants will play an important role in the software industry. However, current software development assistants tend to be unreliable, often producing incorrect, unsafe, or low-quality code. We seek to resolve these issues by introducing a holistic architecture for constructing, training, and using trustworthy AI software development assistants. In the center of the architecture, there is a foundational LLM trained on datasets representative of real-world coding scenarios and complex software architectures, and fine-tuned on code quality criteria beyond correctness. The LLM will make use of graph-based code representations for advanced semantic comprehension. We envision a knowledge graph integrated into the system to provide up-to-date background knowledge and to enable the assistant to provide appropriate explanations. Finally, a modular framework for constrained decoding will ensure that certain guarantees (e.g., for correctness and security) hold for the generated code.

Studies on interpersonal conflict have a long history and contain many suggestions for conflict typology. We use this as the basis of a novel annotation scheme and release a new dataset of situations and conflict aspect annotations. We then build a classifier to predict whether someone will perceive the actions of one individual as right or wrong in a given situation. Our analyses include conflict aspects, but also generated clusters, which are human validated, and show differences in conflict content based on the relationship of participants to the author. Our findings have important implications for understanding conflict and social norms.

Distributed control systems (DCS) are essential to operate complex industrial processes. A major part of a DCS is the alarm system, which helps plant operators to keep the processes stable and safe. Alarms are defined as threshold values on individual signals taking into account minimum reaction time of the human operator. In reality, however, alarms are often noisy and overwhelming, and thus can be easily overlooked by the operators. Early alarm prediction can give the operator more time to react and introduce corrective actions to avoid downtime and negative impact on human safety and the environment. In this context, we introduce Alarm Prediction Transformer (APT), a multimodal Transformer-based machine learning model for early alarm prediction based on the combination of recent events and signal data. Specifically, we propose two novel fusion strategies and three methods of label encoding with various levels of granularity. Given a window of several minutes of event logs and signal data, our model predicts whether an alarm is going to be triggered after a few minutes and, if yes, it also predicts its location. Our experiments on two novel real industrial plant data sets and a simulated data set show that the model is capable of predicting alarms with the given horizon and that our proposed fusion technique combining inputs from different modalities, i. e. events and signals, yields more accurate results than any of the modalities alone or conventional fusion techniques. • Multimodal Transformers can successfully predict alarms in industrial settings. • Transformers can predict future alarms based on recent events and signal data. • The proposed hybrid fusion performs on par with or better than early or late fusion. • Granular alarm encoding yields more accurate prediction than the whole tag.

Recent language modeling performance has been greatly improved by the use of external memory. This memory encodes the context so that similar contexts can be recalled during decoding. This similarity depends on how the model learns to encode context, which can be altered to include other attributes, such as style. We construct and evaluate an architecture for this purpose, using corpora annotated for politeness, formality, and toxicity. Through extensive experiments and human evaluation we demonstrate the potential of our method to generate text while controlling style. We find that style-specific datastores improve generation performance, though results vary greatly across styles, and the effect of pretraining data and specific styles should be explored in future work.

Instead of using a single ground truth for language processing tasks, several recent studies have examined how to represent and predict the labels of the set of annotators. However, often little or no information about annotators is known, or the set of annotators is small. In this work, we examine a corpus of social media posts about conflict from a set of 13k annotators and 210k judgements of social norms. We provide a novel experimental setup that applies personalization methods to the modeling of annotators and compare their effectiveness for predicting the perception of social norms. We further provide an analysis of performance across subsets of social situations that vary by the closeness of the relationship between parties in conflict, and assess where personalization helps the most.

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.