Constantin A. Rothkopf

Katharina Hämmerl, Björn Deiseroth, Patrick Schramowski et al.

Pre-trained multilingual language models (PMLMs) are commonly used when dealing with data from multiple languages and cross-lingual transfer. However, PMLMs are trained on varying amounts of data for each language. In practice this means their performance is often much better on English than many other languages. We explore to what extent this also applies to moral norms. Do the models capture moral norms from English and impose them on other languages? Do the models exhibit random and thus potentially harmful beliefs in certain languages? Both these issues could negatively impact cross-lingual transfer and potentially lead to harmful outcomes. In this paper, we (1) apply the MoralDirection framework to multilingual models, comparing results in German, Czech, Arabic, Chinese, and English, (2) analyse model behaviour on filtered parallel subtitles corpora, and (3) apply the models to a Moral Foundations Questionnaire, comparing with human responses from different countries. Our experiments demonstrate that, indeed, PMLMs encode differing moral biases, but these do not necessarily correspond to cultural differences or commonalities in human opinions. We release our code and models.

Matthias Schultheis, Constantin A. Rothkopf, Heinz Koeppl

Commonly in reinforcement learning (RL), rewards are discounted over time using an exponential function to model time preference, thereby bounding the expected long-term reward. In contrast, in economics and psychology, it has been shown that humans often adopt a hyperbolic discounting scheme, which is optimal when a specific task termination time distribution is assumed. In this work, we propose a theory for continuous-time model-based reinforcement learning generalized to arbitrary discount functions. This formulation covers the case in which there is a non-exponential random termination time. We derive a Hamilton-Jacobi-Bellman (HJB) equation characterizing the optimal policy and describe how it can be solved using a collocation method, which uses deep learning for function approximation. Further, we show how the inverse RL problem can be approached, in which one tries to recover properties of the discount function given decision data. We validate the applicability of our proposed approach on two simulated problems. Our approach opens the way for the analysis of human discounting in sequential decision-making tasks.

Florian Kadner, Tobias Thomas, David Hoppe et al.

The human prioritization of image regions can be modeled in a time invariant fashion with saliency maps or sequentially with scanpath models. However, while both types of models have steadily improved on several benchmarks and datasets, there is still a considerable gap in predicting human gaze. Here, we leverage two recent developments to reduce this gap: theoretical analyses establishing a principled framework for predicting the next gaze target and the empirical measurement of the human cost for gaze switches independently of image content. We introduce an algorithm in the framework of sequential decision making, which converts any static saliency map into a sequence of dynamic history-dependent value maps, which are recomputed after each gaze shift. These maps are based on 1) a saliency map provided by an arbitrary saliency model, 2) the recently measured human cost function quantifying preferences in magnitude and direction of eye movements, and 3) a sequential exploration bonus, which changes with each subsequent gaze shift. The parameters of the spatial extent and temporal decay of this exploration bonus are estimated from human gaze data. The relative contributions of these three components were optimized on the MIT1003 dataset for the NSS score and are sufficient to significantly outperform predictions of the next gaze target on NSS and AUC scores for five state of the art saliency models on three image data sets. Thus, we provide an implementation of human gaze preferences, which can be used to improve arbitrary saliency models' predictions of humans' next gaze targets.

Dominik Straub, Matthias Schultheis, Heinz Koeppl et al.

Inverse optimal control can be used to characterize behavior in sequential decision-making tasks. Most existing work, however, is limited to fully observable or linear systems, or requires the action signals to be known. Here, we introduce a probabilistic approach to inverse optimal control for partially observable stochastic non-linear systems with unobserved action signals, which unifies previous approaches to inverse optimal control with maximum causal entropy formulations. Using an explicit model of the noise characteristics of the sensory and motor systems of the agent in conjunction with local linearization techniques, we derive an approximate likelihood function for the model parameters, which can be computed within a single forward pass. We present quantitative evaluations on stochastic and partially observable versions of two classic control tasks and two human behavioral tasks. Importantly, we show that our method can disentangle perceptual factors and behavioral costs despite the fact that epistemic and pragmatic actions are intertwined in sequential decision-making under uncertainty, such as in active sensing and active learning. The proposed method has broad applicability, ranging from imitation learning to sensorimotor neuroscience.

Dominik Straub, Tobias F. Niehues, Jan Peters et al.

Bayesian observer and actor models have provided normative explanations for many behavioral phenomena in perception, sensorimotor control, and other areas of cognitive science and neuroscience. They attribute behavioral variability and biases to interpretable entities such as perceptual and motor uncertainty, prior beliefs, and behavioral costs. However, when extending these models to more naturalistic tasks with continuous actions, solving the Bayesian decision-making problem is often analytically intractable. Inverse decision-making, i.e. performing inference over the parameters of such models given behavioral data, is computationally even more difficult. Therefore, researchers typically constrain their models to easily tractable components, such as Gaussian distributions or quadratic cost functions, or resort to numerical approximations. To overcome these limitations, we amortize the Bayesian actor using a neural network trained on a wide range of parameter settings in an unsupervised fashion. Using the pre-trained neural network enables performing efficient gradient-based Bayesian inference of the Bayesian actor model's parameters. We show on synthetic data that the inferred posterior distributions are in close alignment with those obtained using analytical solutions where they exist. Where no analytical solution is available, we recover posterior distributions close to the ground truth. We then show how our method allows for principled model comparison and how it can be used to disentangle factors that may lead to unidentifiabilities between priors and costs. Finally, we apply our method to empirical data from three sensorimotor tasks and compare model fits with different cost functions to show that it can explain individuals' behavioral patterns.

Alina Böhm, Tim Schneider, Boris Belousov et al.

This paper explores active sensing strategies that employ vision-based tactile sensors for robotic perception and classification of fabric textures. We formalize the active sampling problem in the context of tactile fabric recognition and provide an implementation of information-theoretic exploration strategies based on minimizing predictive entropy and variance of probabilistic models. Through ablation studies and human experiments, we investigate which components are crucial for quick and reliable texture recognition. Along with the active sampling strategies, we evaluate neural network architectures, representations of uncertainty, influence of data augmentation, and dataset variability. By evaluating our method on a previously published Active Clothing Perception Dataset and on a real robotic system, we establish that the choice of the active exploration strategy has only a minor influence on the recognition accuracy, whereas data augmentation and dropout rate play a significantly larger role. In a comparison study, while humans achieve 66.9% recognition accuracy, our best approach reaches 90.0% in under 5 touches, highlighting that vision-based tactile sensors are highly effective for fabric texture recognition.

Antonia Wüst, Tim Woydt, Lukas Helff et al.

Recently, newly developed Vision-Language Models (VLMs), such as OpenAI's o1, have emerged, seemingly demonstrating advanced reasoning capabilities across text and image modalities. However, the depth of these advances in language-guided perception and abstract reasoning remains underexplored, and it is unclear whether these models can truly live up to their ambitious promises. To assess the progress and identify shortcomings, we enter the wonderland of Bongard problems, a set of classic visual reasoning puzzles that require human-like abilities of pattern recognition and abstract reasoning. With our extensive evaluation setup, we show that while VLMs occasionally succeed in identifying discriminative concepts and solving some of the problems, they frequently falter. Surprisingly, even elementary concepts that may seem trivial to humans, such as simple spirals, pose significant challenges. Moreover, when explicitly asked to recognize ground truth concepts, they continue to falter, suggesting not only a lack of understanding of these elementary visual concepts but also an inability to generalize to unseen concepts. We compare the results of VLMs to human performance and observe that a significant gap remains between human visual reasoning capabilities and machine cognition.

Quentin Delfosse, Jannis Blüml, Fabian Tatai et al.

Deep reinforcement learning (RL) agents achieve impressive results in a wide variety of tasks, but they lack zero-shot adaptation capabilities. While most robustness evaluations focus on tasks complexifications, for which human also struggle to maintain performances, no evaluation has been performed on tasks simplifications. To tackle this issue, we introduce HackAtari, a set of task variations of the Arcade Learning Environments. We use it to demonstrate that, contrary to humans, RL agents systematically exhibit huge performance drops on simpler versions of their training tasks, uncovering agents' consistent reliance on shortcuts. Our analysis across multiple algorithms and architectures highlights the persistent gap between RL agents and human behavioral intelligence, underscoring the need for new benchmarks and methodologies that enforce systematic generalization testing beyond static evaluation protocols. Training and testing in the same environment is not enough to obtain agents equipped with human-like intelligence.

Tim Woydt, Moritz Willig, Antonia Wüst et al.

Strong meta-learning capabilities for systematic compositionality are emerging as an important skill for navigating the complex and changing tasks of today's world. However, in presenting models for robust adaptation to novel environments, it is important to refrain from making unsupported claims about the performance of meta-learning systems that ultimately do not stand up to scrutiny. While Fodor and Pylyshyn famously posited that neural networks inherently lack this capacity as they are unable to model compositional representations or structure-sensitive operations, and thus are not a viable model of the human mind, Lake and Baroni recently presented meta-learning as a pathway to compositionality. In this position paper, we critically revisit this claim and highlight limitations in the proposed meta-learning framework for compositionality. Our analysis shows that modern neural meta-learning systems can only perform such tasks, if at all, under a very narrow and restricted definition of a meta-learning setup. We therefore claim that `Fodor and Pylyshyn's legacy' persists, and to date, there is no human-like systematic compositionality learned in neural networks.

Niteesh Midlagajni, Constantin A. Rothkopf

Simulating particle dynamics with high fidelity is crucial for solving real-world interaction and control tasks involving liquids in design, graphics, and robotics. Recently, data-driven approaches, particularly those based on graph neural networks (GNNs), have shown progress in tackling such problems. However, these approaches are often limited to learning fluid behavior in static free-fall environments or simple manipulation settings involving primitive objects, often overlooking complex interactions with dynamically moving kinematic rigid bodies. Here, we propose a GNN-based framework designed from the ground up to learn the dynamics of liquids under rigid body interactions and active manipulations, where particles are represented as graph nodes and particle-object collisions are handled using surface representations with the bounding volume hierarchy (BVH) algorithm. This approach enables the network to model complex interactions between liquid particles and intricate surface geometries. Our model accurately captures fluid behavior in dynamic settings and can also function as a simulator in static free-fall environments. Despite being trained on a single-object manipulation task of pouring, our model generalizes effectively to environments with unseen objects and novel manipulation tasks such as stirring and scooping. Finally, we show that the learned dynamics can be leveraged to solve control and manipulation tasks using gradient-based optimization methods.

Nils Neupärtl, Constantin A. Rothkopf

Bayesian models of behavior have provided computational level explanations in a range of psychophysical tasks. One fundamental experimental paradigm is the production or reproduction task, in which subjects are instructed to generate an action that either reproduces a previously sensed stimulus magnitude or achieves a target response. This type of task therefore distinguishes itself from other psychophysical tasks in that the responses are on a continuum and effort plays an important role with increasing response magnitude. Based on Bayesian decision theory we present an inference method to recover perceptual uncertainty, response variability, and the cost function underlying human responses. Crucially, the cost function is parameterized such that effort is explicitly included. We present a hybrid inference method employing MCMC sampling utilizing appropriate proposal distributions and an inner loop utilizing amortized inference with a neural network that approximates the mode of the optimal response distribution. We show how this model can be utilized to avoid unidentifiability of experimental designs and that parameters can be recovered through validation on synthetic and application to experimental data. Our approach will enable behavioral scientists to perform Bayesian inference of decision making parameters in production and reproduction tasks.

Matthias Schultheis, Dominik Straub, Constantin A. Rothkopf

Computational level explanations based on optimal feedback control with signal-dependent noise have been able to account for a vast array of phenomena in human sensorimotor behavior. However, commonly a cost function needs to be assumed for a task and the optimality of human behavior is evaluated by comparing observed and predicted trajectories. Here, we introduce inverse optimal control with signal-dependent noise, which allows inferring the cost function from observed behavior. To do so, we formalize the problem as a partially observable Markov decision process and distinguish between the agent's and the experimenter's inference problems. Specifically, we derive a probabilistic formulation of the evolution of states and belief states and an approximation to the propagation equation in the linear-quadratic Gaussian problem with signal-dependent noise. We extend the model to the case of partial observability of state variables from the point of view of the experimenter. We show the feasibility of the approach through validation on synthetic data and application to experimental data. Our approach enables recovering the costs and benefits implicit in human sequential sensorimotor behavior, thereby reconciling normative and descriptive approaches in a computational framework.

Matej Zečević, Devendra Singh Dhami, Constantin A. Rothkopf et al.

In explanatory interactive learning (XIL) the user queries the learner, then the learner explains its answer to the user and finally the loop repeats. XIL is attractive for two reasons, (1) the learner becomes better and (2) the user's trust increases. For both reasons to hold, the learner's explanations must be useful to the user and the user must be allowed to ask useful questions. Ideally, both questions and explanations should be grounded in a causal model since they avoid spurious fallacies. Ultimately, we seem to seek a causal variant of XIL. The question part on the user's end we believe to be solved since the user's mental model can provide the causal model. But how would the learner provide causal explanations? In this work we show that existing explanation methods are not guaranteed to be causal even when provided with a Structural Causal Model (SCM). Specifically, we use the popular, proclaimed causal explanation method CXPlain to illustrate how the generated explanations leave open the question of truly causal explanations. Thus as a step towards causal XIL, we propose a solution to the lack of causal explanations. We solve this problem by deriving from first principles an explanation method that makes full use of a given SCM, which we refer to as SC$\textbf{E}$ ($\textbf{E}$ standing for explanation). Since SCEs make use of structural information, any causal graph learner can now provide human-readable explanations. We conduct several experiments including a user study with 22 participants to investigate the virtue of SCE as causal explanations of SCMs.

Susanne Trick, Constantin A. Rothkopf

Combining the outputs of multiple classifiers or experts into a single probabilistic classification is a fundamental task in machine learning with broad applications from classifier fusion to expert opinion pooling. Here we present a hierarchical Bayesian model of probabilistic classifier fusion based on a new correlated Dirichlet distribution. This distribution explicitly models positive correlations between marginally Dirichlet-distributed random vectors thereby allowing explicit modeling of correlations between base classifiers or experts. The proposed model naturally accommodates the classic Independent Opinion Pool and other independent fusion algorithms as special cases. It is evaluated by uncertainty reduction and correctness of fusion on synthetic and real-world data sets. We show that a change in performance of the fused classifier due to uncertainty reduction can be Bayes optimal even for highly correlated base classifiers.

Patrick Schramowski, Cigdem Turan, Nico Andersen et al.

Artificial writing is permeating our lives due to recent advances in large-scale, transformer-based language models (LMs) such as BERT, its variants, GPT-2/3, and others. Using them as pre-trained models and fine-tuning them for specific tasks, researchers have extended state of the art for many NLP tasks and shown that they capture not only linguistic knowledge but also retain general knowledge implicitly present in the data. Unfortunately, LMs trained on unfiltered text corpora suffer from degenerated and biased behaviour. While this is well established, we show that recent LMs also contain human-like biases of what is right and wrong to do, some form of ethical and moral norms of the society -- they bring a "moral direction" to surface. That is, we show that these norms can be captured geometrically by a direction, which can be computed, e.g., by a PCA, in the embedding space, reflecting well the agreement of phrases to social norms implicitly expressed in the training texts and providing a path for attenuating or even preventing toxic degeneration in LMs. Being able to rate the (non-)normativity of arbitrary phrases without explicitly training the LM for this task, we demonstrate the capabilities of the "moral direction" for guiding (even other) LMs towards producing normative text and showcase it on RealToxicityPrompts testbed, preventing the neural toxic degeneration in GPT-2.

Cigdem Turan, Patrick Schramowski, Constantin Rothkopf et al.

This work introduces Alfie, an interactive robot that is capable of answering moral (deontological) questions of a user. The interaction of Alfie is designed in a way in which the user can offer an alternative answer when the user disagrees with the given answer so that Alfie can learn from its interactions. Alfie's answers are based on a sentence embedding model that uses state-of-the-art language models, e.g. Universal Sentence Encoder and BERT. Alfie is implemented on a Furhat Robot, which provides a customizable user interface to design a social robot.

Patrick Schramowski, Cigdem Turan, Sophie Jentzsch et al.

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? Jentzsch et al.(2019) showed that applying machine learning to human texts can extract deontological ethical reasoning about "right" and "wrong" conduct by calculating a moral bias score on a sentence level using sentence embeddings. The machine learned that 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. However, the evaluated moral bias was restricted to simple actions -- one verb -- and a ranking of actions with surrounding context. Recently BERT ---and variants such as RoBERTa and SBERT--- has set a new state-of-the-art performance for a wide range of NLP tasks. But has BERT also a better moral compass? In this paper, we discuss and show that this is indeed the case. Thus, recent improvements of language representations also improve the representation of the underlying ethical and moral values of the machine. We argue that through an advanced semantic representation of text, BERT allows one to get better insights of moral and ethical values implicitly represented in text. This enables the Moral Choice Machine (MCM) to extract more accurate imprints of moral choices and ethical values.

Nikita Araslanov, Constantin Rothkopf, Stefan Roth

Most approaches to visual scene analysis have emphasised parallel processing of the image elements. However, one area in which the sequential nature of vision is apparent, is that of segmenting multiple, potentially similar and partially occluded objects in a scene. In this work, we revisit the recurrent formulation of this challenging problem in the context of reinforcement learning. Motivated by the limitations of the global max-matching assignment of the ground-truth segments to the recurrent states, we develop an actor-critic approach in which the actor recurrently predicts one instance mask at a time and utilises the gradient from a concurrently trained critic network. We formulate the state, action, and the reward such as to let the critic model long-term effects of the current prediction and incorporate this information into the gradient signal. Furthermore, to enable effective exploration in the inherently high-dimensional action space of instance masks, we learn a compact representation using a conditional variational auto-encoder. We show that our actor-critic model consistently provides accuracy benefits over the recurrent baseline on standard instance segmentation benchmarks.

Kristian Kersting, Jan Peters, Constantin Rothkopf

The Federal Government of Germany aims to boost the research in the field of Artificial Intelligence (AI). For instance, 100 new professorships are said to be established. However, the white paper of the government does not answer what an AI professorship is at all. In order to give colleagues, politicians, and citizens an idea, we present a view that is often followed when appointing professors for AI at German and international universities. We hope that it will help to establish a guideline with internationally accepted measures and thus make the public debate more informed.

Stefan Depeweg, Constantin A. Rothkopf, Frank Jäkel

More than 50 years ago Bongard introduced 100 visual concept learning problems as a testbed for intelligent vision systems. These problems are now known as Bongard problems. Although they are well known in the cognitive science and AI communities only moderate progress has been made towards building systems that can solve a substantial subset of them. In the system presented here, visual features are extracted through image processing and then translated into a symbolic visual vocabulary. We introduce a formal language that allows representing complex visual concepts based on this vocabulary. Using this language and Bayesian inference, complex visual concepts can be induced from the examples that are provided in each Bongard problem. Contrary to other concept learning problems the examples from which concepts are induced are not random in Bongard problems, instead they are carefully chosen to communicate the concept, hence requiring pragmatic reasoning. Taking pragmatic reasoning into account we find good agreement between the concepts with high posterior probability and the solutions formulated by Bongard himself. While this approach is far from solving all Bongard problems, it solves the biggest fraction yet.

V S R Veeravasarapu, Constantin Rothkopf, Ramesh Visvanathan

Generalization performance of trained computer vision systems that use computer graphics (CG) generated data is not yet effective due to the concept of 'domain-shift' between virtual and real data. Although simulated data augmented with a few real world samples has been shown to mitigate domain shift and improve transferability of trained models, guiding or bootstrapping the virtual data generation with the distributions learnt from target real world domain is desired, especially in the fields where annotating even few real images is laborious (such as semantic labeling, and intrinsic images etc.). In order to address this problem in an unsupervised manner, our work combines recent advances in CG (which aims to generate stochastic scene layouts coupled with large collections of 3D object models) and generative adversarial training (which aims train generative models by measuring discrepancy between generated and real data in terms of their separability in the space of a deep discriminatively-trained classifier). Our method uses iterative estimation of the posterior density of prior distributions for a generative graphical model. This is done within a rejection sampling framework. Initially, we assume uniform distributions as priors on the parameters of a scene described by a generative graphical model. As iterations proceed the prior distributions get updated to distributions that are closer to the (unknown) distributions of target data. We demonstrate the utility of adversarially tuned scene generation on two real-world benchmark datasets (CityScapes and CamVid) for traffic scene semantic labeling with a deep convolutional net (DeepLab). We realized performance improvements by 2.28 and 3.14 points (using the IoU metric) between the DeepLab models trained on simulated sets prepared from the scene generation models before and after tuning to CityScapes and CamVid respectively.

V S R Veeravasarapu, Constantin Rothkopf, Visvanathan Ramesh

The use of simulated virtual environments to train deep convolutional neural networks (CNN) is a currently active practice to reduce the (real)data-hungriness of the deep CNN models, especially in application domains in which large scale real data and/or groundtruth acquisition is difficult or laborious. Recent approaches have attempted to harness the capabilities of existing video games, animated movies to provide training data with high precision groundtruth. However, a stumbling block is in how one can certify generalization of the learned models and their usefulness in real world data sets. This opens up fundamental questions such as: What is the role of photorealism of graphics simulations in training CNN models? Are the trained models valid in reality? What are possible ways to reduce the performance bias? In this work, we begin to address theses issues systematically in the context of urban semantic understanding with CNNs. Towards this end, we (a) propose a simple probabilistic urban scene model, (b) develop a parametric rendering tool to synthesize the data with groundtruth, followed by (c) a systematic exploration of the impact of level-of-realism on the generality of the trained CNN model to real world; and domain adaptation concepts to minimize the performance bias.

V S R Veeravasarapu, Rudra Narayan Hota, Constantin Rothkopf et al.

Rapid advances in computation, combined with latest advances in computer graphics simulations have facilitated the development of vision systems and training them in virtual environments. One major stumbling block is in certification of the designs and tuned parameters of these systems to work in real world. In this paper, we begin to explore the fundamental question: Which type of information transfer is more analogous to real world? Inspired from the performance characterization methodology outlined in the 90's, we note that insights derived from simulations can be qualitative or quantitative depending on the degree of the fidelity of models used in simulations and the nature of the questions posed by the experimenter. We adapt the methodology in the context of current graphics simulation tools for modeling data generation processes and, for systematic performance characterization and trade-off analysis for vision system design leading to qualitative and quantitative insights. In concrete, we examine invariance assumptions used in vision algorithms for video surveillance settings as a case study and assess the degree to which those invariance assumptions deviate as a function of contextual variables on both graphics simulations and in real data. As computer graphics rendering quality improves, we believe teasing apart the degree to which model assumptions are valid via systematic graphics simulation can be a significant aid to assisting more principled ways of approaching vision system design and performance modeling.

V S R Veeravasarapu, Rudra Narayan Hota, Constantin Rothkopf et al.

As the computer vision matures into a systems science and engineering discipline, there is a trend in leveraging latest advances in computer graphics simulations for performance evaluation, learning, and inference. However, there is an open question on the utility of graphics simulations for vision with apparently contradicting views in the literature. In this paper, we place the results from the recent literature in the context of performance characterization methodology outlined in the 90's and note that insights derived from simulations can be qualitative or quantitative depending on the degree of fidelity of models used in simulation and the nature of the question posed by the experimenter. We describe a simulation platform that incorporates latest graphics advances and use it for systematic performance characterization and trade-off analysis for vision system design. We verify the utility of the platform in a case study of validating a generative model inspired vision hypothesis, Rank-Order consistency model, in the contexts of global and local illumination changes, and bad weather, and high-frequency noise. Our approach establishes the link between alternative viewpoints, involving models with physics based semantics and signal and perturbation semantics and confirms insights in literature on robust change detection.