Eyke Hüllermeier

Eli Verwimp, Rahaf Aljundi, Shai Ben-David et al. · deepmind

Continual learning is a subfield of machine learning, which aims to allow machine learning models to continuously learn on new data, by accumulating knowledge without forgetting what was learned in the past. In this work, we take a step back, and ask: "Why should one care about continual learning in the first place?". We set the stage by examining recent continual learning papers published at four major machine learning conferences, and show that memory-constrained settings dominate the field. Then, we discuss five open problems in machine learning, and even though they might seem unrelated to continual learning at first sight, we show that continual learning will inevitably be part of their solution. These problems are model editing, personalization and specialization, on-device learning, faster (re-)training and reinforcement learning. Finally, by comparing the desiderata from these unsolved problems and the current assumptions in continual learning, we highlight and discuss four future directions for continual learning research. We hope that this work offers an interesting perspective on the future of continual learning, while displaying its potential value and the paths we have to pursue in order to make it successful. This work is the result of the many discussions the authors had at the Dagstuhl seminar on Deep Continual Learning, in March 2023.

Viktor Bengs, Eyke Hüllermeier, Willem Waegeman

Uncertainty quantification has received increasing attention in machine learning in the recent past. In particular, a distinction between aleatoric and epistemic uncertainty has been found useful in this regard. The latter refers to the learner's (lack of) knowledge and appears to be especially difficult to measure and quantify. In this paper, we analyse a recent proposal based on the idea of a second-order learner, which yields predictions in the form of distributions over probability distributions. While standard (first-order) learners can be trained to predict accurate probabilities, namely by minimising suitable loss functions on sample data, we show that loss minimisation does not work for second-order predictors: The loss functions proposed for inducing such predictors do not incentivise the learner to represent its epistemic uncertainty in a faithful way.

Fabian Fumagalli, Maximilian Muschalik, Patrick Kolpaczki et al.

Predominately in explainable artificial intelligence (XAI) research, the Shapley value (SV) is applied to determine feature attributions for any black box model. Shapley interaction indices extend the SV to define any-order feature interactions. Defining a unique Shapley interaction index is an open research question and, so far, three definitions have been proposed, which differ by their choice of axioms. Moreover, each definition requires a specific approximation technique. Here, we propose SHAPley Interaction Quantification (SHAP-IQ), an efficient sampling-based approximator to compute Shapley interactions for arbitrary cardinal interaction indices (CII), i.e. interaction indices that satisfy the linearity, symmetry and dummy axiom. SHAP-IQ is based on a novel representation and, in contrast to existing methods, we provide theoretical guarantees for its approximation quality, as well as estimates for the variance of the point estimates. For the special case of SV, our approach reveals a novel representation of the SV and corresponds to Unbiased KernelSHAP with a greatly simplified calculation. We illustrate the computational efficiency and effectiveness by explaining language, image classification and high-dimensional synthetic models.

Fabian Fumagalli, Maximilian Muschalik, Eyke Hüllermeier et al.

Explainable Artificial Intelligence (XAI) has mainly focused on static learning scenarios so far. We are interested in dynamic scenarios where data is sampled progressively, and learning is done in an incremental rather than a batch mode. We seek efficient incremental algorithms for computing feature importance (FI) measures, specifically, an incremental FI measure based on feature marginalization of absent features similar to permutation feature importance (PFI). We propose an efficient, model-agnostic algorithm called iPFI to estimate this measure incrementally and under dynamic modeling conditions including concept drift. We prove theoretical guarantees on the approximation quality in terms of expectation and variance. To validate our theoretical findings and the efficacy of our approaches compared to traditional batch PFI, we conduct multiple experimental studies on benchmark data with and without concept drift.

Yusuf Sale, Michele Caprio, Eyke Hüllermeier

Adequate uncertainty representation and quantification have become imperative in various scientific disciplines, especially in machine learning and artificial intelligence. As an alternative to representing uncertainty via one single probability measure, we consider credal sets (convex sets of probability measures). The geometric representation of credal sets as $d$-dimensional polytopes implies a geometric intuition about (epistemic) uncertainty. In this paper, we show that the volume of the geometric representation of a credal set is a meaningful measure of epistemic uncertainty in the case of binary classification, but less so for multi-class classification. Our theoretical findings highlight the crucial role of specifying and employing uncertainty measures in machine learning in an appropriate way, and for being aware of possible pitfalls.

Viktor Bengs, Eyke Hüllermeier, Willem Waegeman

It is well known that accurate probabilistic predictors can be trained through empirical risk minimisation with proper scoring rules as loss functions. While such learners capture so-called aleatoric uncertainty of predictions, various machine learning methods have recently been developed with the goal to let the learner also represent its epistemic uncertainty, i.e., the uncertainty caused by a lack of knowledge and data. An emerging branch of the literature proposes the use of a second-order learner that provides predictions in terms of distributions on probability distributions. However, recent work has revealed serious theoretical shortcomings for second-order predictors based on loss minimisation. In this paper, we generalise these findings and prove a more fundamental result: There seems to be no loss function that provides an incentive for a second-order learner to faithfully represent its epistemic uncertainty in the same manner as proper scoring rules do for standard (first-order) learners. As a main mathematical tool to prove this result, we introduce the generalised notion of second-order scoring rules.

Michele Caprio, Yusuf Sale, Eyke Hüllermeier et al.

In their seminal 1990 paper, Wasserman and Kadane establish an upper bound for the Bayes' posterior probability of a measurable set $A$, when the prior lies in a class of probability measures $\mathcal{P}$ and the likelihood is precise. They also give a sufficient condition for such upper bound to hold with equality. In this paper, we introduce a generalization of their result by additionally addressing uncertainty related to the likelihood. We give an upper bound for the posterior probability when both the prior and the likelihood belong to a set of probabilities. Furthermore, we give a sufficient condition for this upper bound to become an equality. This result is interesting on its own, and has the potential of being applied to various fields of engineering (e.g. model predictive control), machine learning, and artificial intelligence.

Maximilian Muschalik, Fabian Fumagalli, Rohit Jagtani et al.

Post-hoc explanation techniques such as the well-established partial dependence plot (PDP), which investigates feature dependencies, are used in explainable artificial intelligence (XAI) to understand black-box machine learning models. While many real-world applications require dynamic models that constantly adapt over time and react to changes in the underlying distribution, XAI, so far, has primarily considered static learning environments, where models are trained in a batch mode and remain unchanged. We thus propose a novel model-agnostic XAI framework called incremental PDP (iPDP) that extends on the PDP to extract time-dependent feature effects in non-stationary learning environments. We formally analyze iPDP and show that it approximates a time-dependent variant of the PDP that properly reacts to real and virtual concept drift. The time-sensitivity of iPDP is controlled by a single smoothing parameter, which directly corresponds to the variance and the approximation error of iPDP in a static learning environment. We illustrate the efficacy of iPDP by showcasing an example application for drift detection and conducting multiple experiments on real-world and synthetic data sets and streams.

Alireza Javanmardi, Eyke Hüllermeier

The main objective of Prognostics and Health Management is to estimate the Remaining Useful Lifetime (RUL), namely, the time that a system or a piece of equipment is still in working order before starting to function incorrectly. In recent years, numerous machine learning algorithms have been proposed for RUL estimation, mainly focusing on providing more accurate RUL predictions. However, there are many sources of uncertainty in the problem, such as inherent randomness of systems failure, lack of knowledge regarding their future states, and inaccuracy of the underlying predictive models, making it infeasible to predict the RULs precisely. Hence, it is of utmost importance to quantify the uncertainty alongside the RUL predictions. In this work, we investigate the conformal prediction (CP) framework that represents uncertainty by predicting sets of possible values for the target variable (intervals in the case of RUL) instead of making point predictions. Under very mild technical assumptions, CP formally guarantees that the actual value (true RUL) is covered by the predicted set with a degree of certainty that can be prespecified. We study three CP algorithms to conformalize any single-point RUL predictor and turn it into a valid interval predictor. Finally, we conformalize two single-point RUL predictors, deep convolutional neural networks and gradient boosting, and illustrate their performance on the Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) data sets.

Subhabrata Dutta, Timo Kaufmann, Goran Glavaš et al.

While there is a widespread belief that artificial general intelligence (AGI) -- or even superhuman AI -- is imminent, complex problems in expert domains are far from being solved. We argue that such problems require human-AI cooperation and that the current state of the art in generative AI is unable to play the role of a reliable partner due to a multitude of shortcomings, including difficulty to keep track of a complex solution artifact (e.g., a software program), limited support for versatile human preference expression and lack of adapting to human preference in an interactive setting. To address these challenges, we propose HAICo2, a novel human-AI co-construction framework. We take first steps towards a formalization of HAICo2 and discuss the difficult open research problems that it faces.

Thomas Mortier, Viktor Bengs, Eyke Hüllermeier et al.

Multi-class classification methods that produce sets of probabilistic classifiers, such as ensemble learning methods, are able to model aleatoric and epistemic uncertainty. Aleatoric uncertainty is then typically quantified via the Bayes error, and epistemic uncertainty via the size of the set. In this paper, we extend the notion of calibration, which is commonly used to evaluate the validity of the aleatoric uncertainty representation of a single probabilistic classifier, to assess the validity of an epistemic uncertainty representation obtained by sets of probabilistic classifiers. Broadly speaking, we call a set of probabilistic classifiers calibrated if one can find a calibrated convex combination of these classifiers. To evaluate this notion of calibration, we propose a novel nonparametric calibration test that generalizes an existing test for single probabilistic classifiers to the case of sets of probabilistic classifiers. Making use of this test, we empirically show that ensembles of deep neural networks are often not well calibrated.

Maximilian Muschalik, Fabian Fumagalli, Barbara Hammer et al.

Existing methods for explainable artificial intelligence (XAI), including popular feature importance measures such as SAGE, are mostly restricted to the batch learning scenario. However, machine learning is often applied in dynamic environments, where data arrives continuously and learning must be done in an online manner. Therefore, we propose iSAGE, a time- and memory-efficient incrementalization of SAGE, which is able to react to changes in the model as well as to drift in the data-generating process. We further provide efficient feature removal methods that break (interventional) and retain (observational) feature dependencies. Moreover, we formally analyze our explanation method to show that iSAGE adheres to similar theoretical properties as SAGE. Finally, we evaluate our approach in a thorough experimental analysis based on well-established data sets and data streams with concept drift.

Alireza Javanmardi, Yusuf Sale, Paul Hofman et al.

While the predictions produced by conformal prediction are set-valued, the data used for training and calibration is supposed to be precise. In the setting of superset learning or learning from partial labels, a variant of weakly supervised learning, it is exactly the other way around: training data is possibly imprecise (set-valued), but the model induced from this data yields precise predictions. In this paper, we combine the two settings by making conformal prediction amenable to set-valued training data. We propose a generalization of the conformal prediction procedure that can be applied to set-valued training and calibration data. We prove the validity of the proposed method and present experimental studies in which it compares favorably to natural baselines.

Thomas Mortier, Eyke Hüllermeier, Krzysztof Dembczyński et al.

Set-valued prediction is a well-known concept in multi-class classification. When a classifier is uncertain about the class label for a test instance, it can predict a set of classes instead of a single class. In this paper, we focus on hierarchical multi-class classification problems, where valid sets (typically) correspond to internal nodes of the hierarchy. We argue that this is a very strong restriction, and we propose a relaxation by introducing the notion of representation complexity for a predicted set. In combination with probabilistic classifiers, this leads to a challenging inference problem for which specific combinatorial optimization algorithms are needed. We propose three methods and evaluate them on benchmark datasets: a naïve approach that is based on matrix-vector multiplication, a reformulation as a knapsack problem with conflict graph, and a recursive tree search method. Experimental results demonstrate that the last method is computationally more efficient than the other two approaches, due to a hierarchical factorization of the conditional class distribution.

Amirhossein Vahidi, Lisa Wimmer, Hüseyin Anil Gündüz et al.

Ensembling a neural network is a widely recognized approach to enhance model performance, estimate uncertainty, and improve robustness in deep supervised learning. However, deep ensembles often come with high computational costs and memory demands. In addition, the efficiency of a deep ensemble is related to diversity among the ensemble members which is challenging for large, over-parameterized deep neural networks. Moreover, ensemble learning has not yet seen such widespread adoption, and it remains a challenging endeavor for self-supervised or unsupervised representation learning. Motivated by these challenges, we present a novel self-supervised training regime that leverages an ensemble of independent sub-networks, complemented by a new loss function designed to encourage diversity. Our method efficiently builds a sub-model ensemble with high diversity, leading to well-calibrated estimates of model uncertainty, all achieved with minimal computational overhead compared to traditional deep self-supervised ensembles. To evaluate the effectiveness of our approach, we conducted extensive experiments across various tasks, including in-distribution generalization, out-of-distribution detection, dataset corruption, and semi-supervised settings. The results demonstrate that our method significantly improves prediction reliability. Our approach not only achieves excellent accuracy but also enhances calibration, surpassing baseline performance across a wide range of self-supervised architectures in computer vision, natural language processing, and genomics data.

Mohamed Karim Belaid, Dorra El Mekki, Maximilian Rabus et al.

With the rapid growth of data availability and usage, quantifying the added value of each training data point has become a crucial process in the field of artificial intelligence. The Shapley values have been recognized as an effective method for data valuation, enabling efficient training set summarization, acquisition, and outlier removal. In this paper, we introduce "STI-KNN", an innovative algorithm that calculates the exact pair-interaction Shapley values for KNN models in O(t n^2) time, which is a significant improvement over the O(2^n)$ time complexity of baseline methods. By using STI-KNN, we can efficiently and accurately evaluate the value of individual data points, leading to improved training outcomes and ultimately enhancing the effectiveness of artificial intelligence applications.

Amirhossein Vahidi, Simon Schoßer, Lisa Wimmer et al.

In this paper, we propose a novel probabilistic self-supervised learning via Scoring Rule Minimization (ProSMIN), which leverages the power of probabilistic models to enhance representation quality and mitigate collapsing representations. Our proposed approach involves two neural networks; the online network and the target network, which collaborate and learn the diverse distribution of representations from each other through knowledge distillation. By presenting the input samples in two augmented formats, the online network is trained to predict the target network representation of the same sample under a different augmented view. The two networks are trained via our new loss function based on proper scoring rules. We provide a theoretical justification for ProSMIN's convergence, demonstrating the strict propriety of its modified scoring rule. This insight validates the method's optimization process and contributes to its robustness and effectiveness in improving representation quality. We evaluate our probabilistic model on various downstream tasks, such as in-distribution generalization, out-of-distribution detection, dataset corruption, low-shot learning, and transfer learning. Our method achieves superior accuracy and calibration, surpassing the self-supervised baseline in a wide range of experiments on large-scale datasets like ImageNet-O and ImageNet-C, ProSMIN demonstrates its scalability and real-world applicability.

Jonas Hanselle, Stefan Heid, Johannes Fürnkranz et al.

A scoring system is a simple decision model that checks a set of features, adds a certain number of points to a total score for each feature that is satisfied, and finally makes a decision by comparing the total score to a threshold. Scoring systems have a long history of active use in safety-critical domains such as healthcare and justice, where they provide guidance for making objective and accurate decisions. Given their genuine interpretability, the idea of learning scoring systems from data is obviously appealing from the perspective of explainable AI. In this paper, we propose a practically motivated extension of scoring systems called probabilistic scoring lists (PSL), as well as a method for learning PSLs from data. Instead of making a deterministic decision, a PSL represents uncertainty in the form of probability distributions, or, more generally, probability intervals. Moreover, in the spirit of decision lists, a PSL evaluates features one by one and stops as soon as a decision can be made with enough confidence. To evaluate our approach, we conduct a case study in the medical domain.

Duc Anh Nguyen, Ron Levie, Julian Lienen et al.

The notion of neural collapse refers to several emergent phenomena that have been empirically observed across various canonical classification problems. During the terminal phase of training a deep neural network, the feature embedding of all examples of the same class tend to collapse to a single representation, and the features of different classes tend to separate as much as possible. Neural collapse is often studied through a simplified model, called the unconstrained feature representation, in which the model is assumed to have "infinite expressivity" and can map each data point to any arbitrary representation. In this work, we propose a more realistic variant of the unconstrained feature representation that takes the limited expressivity of the network into account. Empirical evidence suggests that the memorization of noisy data points leads to a degradation (dilation) of the neural collapse. Using a model of the memorization-dilation (M-D) phenomenon, we show one mechanism by which different losses lead to different performances of the trained network on noisy data. Our proofs reveal why label smoothing, a modification of cross-entropy empirically observed to produce a regularization effect, leads to improved generalization in classification tasks.

Clemens Damke, Eyke Hüllermeier

In this work, we study the influence of domain-specific characteristics when defining a meaningful notion of predictive uncertainty on graph data. Previously, the so-called Graph Posterior Network (GPN) model has been proposed to quantify uncertainty in node classification tasks. Given a graph, it uses Normalizing Flows (NFs) to estimate class densities for each node independently and converts those densities into Dirichlet pseudo-counts, which are then dispersed through the graph using the personalized Page-Rank algorithm. The architecture of GPNs is motivated by a set of three axioms on the properties of its uncertainty estimates. We show that those axioms are not always satisfied in practice and therefore propose the family of Committe-based Uncertainty Quantification Graph Neural Networks (CUQ-GNNs), which combine standard Graph Neural Networks with the NF-based uncertainty estimation of Posterior Networks (PostNets). This approach adapts more flexibly to domain-specific demands on the properties of uncertainty estimates. We compare CUQ-GNN against GPN and other uncertainty quantification approaches on common node classification benchmarks and show that it is effective at producing useful uncertainty estimates.

Viktor Bengs, Björn Haddenhorst, Eyke Hüllermeier

We consider the task of identifying the Copeland winner(s) in a dueling bandits problem with ternary feedback. This is an underexplored but practically relevant variant of the conventional dueling bandits problem, in which, in addition to strict preference between two arms, one may observe feedback in the form of an indifference. We provide a lower bound on the sample complexity for any learning algorithm finding the Copeland winner(s) with a fixed error probability. Moreover, we propose POCOWISTA, an algorithm with a sample complexity that almost matches this lower bound, and which shows excellent empirical performance, even for the conventional dueling bandits problem. For the case where the preference probabilities satisfy a specific type of stochastic transitivity, we provide a refined version with an improved worst case sample complexity.

Sascha Henzgen, Eyke Hüllermeier

Measures of rank correlation are commonly used in statistics to capture the degree of concordance between two orderings of the same set of items. Standard measures like Kendall's tau and Spearman's rho coefficient put equal emphasis on each position of a ranking. Yet, motivated by applications in which some of the positions (typically those on the top) are more important than others, a few weighted variants of these measures have been proposed. Most of these generalizations fail to meet desirable formal properties, however. Besides, they are often quite inflexible in the sense of committing to a fixed weighing scheme. In this paper, we propose a weighted rank correlation measure on the basis of fuzzy order relations. Our measure, called scaled gamma, is related to Goodman and Kruskal's gamma rank correlation. It is parametrized by a fuzzy equivalence relation on the rank positions, which in turn is specified conveniently by a so-called scaling function. This approach combines soundness with flexibility: it has a sound formal foundation and allows for weighing rank positions in a flexible way.

Mohamed Karim Belaid, Eyke Hüllermeier, Maximilian Rabus et al.

In recent years, Explainable AI (xAI) attracted a lot of attention as various countries turned explanations into a legal right. xAI allows for improving models beyond the accuracy metric by, e.g., debugging the learned pattern and demystifying the AI's behavior. The widespread use of xAI brought new challenges. On the one hand, the number of published xAI algorithms underwent a boom, and it became difficult for practitioners to select the right tool. On the other hand, some experiments did highlight how easy data scientists could misuse xAI algorithms and misinterpret their results. To tackle the issue of comparing and correctly using feature importance xAI algorithms, we propose Compare-xAI, a benchmark that unifies all exclusive functional testing methods applied to xAI algorithms. We propose a selection protocol to shortlist non-redundant functional tests from the literature, i.e., each targeting a specific end-user requirement in explaining a model. The benchmark encapsulates the complexity of evaluating xAI methods into a hierarchical scoring of three levels, namely, targeting three end-user groups: researchers, practitioners, and laymen in xAI. The most detailed level provides one score per test. The second level regroups tests into five categories (fidelity, fragility, stability, simplicity, and stress tests). The last level is the aggregated comprehensibility score, which encapsulates the ease of correctly interpreting the algorithm's output in one easy to compare value. Compare-xAI's interactive user interface helps mitigate errors in interpreting xAI results by quickly listing the recommended xAI solutions for each ML task and their current limitations. The benchmark is made available at https://karim-53.github.io/cxai/

Jasmin Brandt, Elias Schede, Viktor Bengs et al.

We study the algorithm configuration (AC) problem, in which one seeks to find an optimal parameter configuration of a given target algorithm in an automated way. Recently, there has been significant progress in designing AC approaches that satisfy strong theoretical guarantees. However, a significant gap still remains between the practical performance of these approaches and state-of-the-art heuristic methods. To this end, we introduce AC-Band, a general approach for the AC problem based on multi-armed bandits that provides theoretical guarantees while exhibiting strong practical performance. We show that AC-Band requires significantly less computation time than other AC approaches providing theoretical guarantees while still yielding high-quality configurations.

Julian Lienen, Caglar Demir, Eyke Hüllermeier

In semi-supervised learning, the paradigm of self-training refers to the idea of learning from pseudo-labels suggested by the learner itself. Across various domains, corresponding methods have proven effective and achieve state-of-the-art performance. However, pseudo-labels typically stem from ad-hoc heuristics, relying on the quality of the predictions though without guaranteeing their validity. One such method, so-called credal self-supervised learning, maintains pseudo-supervision in the form of sets of (instead of single) probability distributions over labels, thereby allowing for a flexible yet uncertainty-aware labeling. Again, however, there is no justification beyond empirical effectiveness. To address this deficiency, we make use of conformal prediction, an approach that comes with guarantees on the validity of set-valued predictions. As a result, the construction of credal sets of labels is supported by a rigorous theoretical foundation, leading to better calibrated and less error-prone supervision for unlabeled data. Along with this, we present effective algorithms for learning from credal self-supervision. An empirical study demonstrates excellent calibration properties of the pseudo-supervision, as well as the competitiveness of our method on several benchmark datasets.

Svenja Uhlemeyer, Julian Lienen, Eyke Hüllermeier et al.

For open world applications, deep neural networks (DNNs) need to be aware of previously unseen data and adaptable to evolving environments. Furthermore, it is desirable to detect and learn novel classes which are not included in the DNNs underlying set of semantic classes in an unsupervised fashion. The method proposed in this article builds upon anomaly detection to retrieve out-of-distribution (OoD) data as candidates for new classes. We thereafter extend the DNN by $k$ empty classes and fine-tune it on the OoD data samples. To this end, we introduce two loss functions, which 1) entice the DNN to assign OoD samples to the empty classes and 2) to minimize the inner-class feature distances between them. Thus, instead of ground truth which contains labels for the different novel classes, the DNN obtains a single OoD label together with a distance matrix, which is computed in advance. We perform several experiments for image classification and semantic segmentation, which demonstrate that a DNN can extend its own semantic space by multiple classes without having access to ground truth.

Jasmin Brandt, Marcel Wever, Dimitrios Iliadis et al.

Hyperparameter optimization (HPO) is concerned with the automated search for the most appropriate hyperparameter configuration (HPC) of a parameterized machine learning algorithm. A state-of-the-art HPO method is Hyperband, which, however, has its own parameters that influence its performance. One of these parameters, the maximal budget, is especially problematic: If chosen too small, the budget needs to be increased in hindsight and, as Hyperband is not incremental by design, the entire algorithm must be re-run. This is not only costly but also comes with a loss of valuable knowledge already accumulated. In this paper, we propose incremental variants of Hyperband that eliminate these drawbacks, and show that these variants satisfy theoretical guarantees qualitatively similar to those for the original Hyperband with the "right" budget. Moreover, we demonstrate their practical utility in experiments with benchmark data sets.

Jiawen Wang, Pritha Gupta, Ivan Habernal et al.

Recent studies demonstrate that Large Language Models (LLMs) are vulnerable to different prompt-based attacks, generating harmful content or sensitive information. Both closed-source and open-source LLMs are underinvestigated for these attacks. This paper studies effective prompt injection attacks against the $\mathbf{14}$ most popular open-source LLMs on five attack benchmarks. Current metrics only consider successful attacks, whereas our proposed Attack Success Probability (ASP) also captures uncertainty in the model's response, reflecting ambiguity in attack feasibility. By comprehensively analyzing the effectiveness of prompt injection attacks, we propose a simple and effective hypnotism attack; results show that this attack causes aligned language models, including Stablelm2, Mistral, Openchat, and Vicuna, to generate objectionable behaviors, achieving around $90$% ASP. They also indicate that our ignore prefix attacks can break all $\mathbf{14}$ open-source LLMs, achieving over $60$% ASP on a multi-categorical dataset. We find that moderately well-known LLMs exhibit higher vulnerability to prompt injection attacks, highlighting the need to raise public awareness and prioritize efficient mitigation strategies.

Timo Kaufmann, Paul Weng, Viktor Bengs et al.

Reinforcement learning from human feedback (RLHF) is a variant of reinforcement learning (RL) that learns from human feedback instead of relying on an engineered reward function. Building on prior work on the related setting of preference-based reinforcement learning (PbRL), it stands at the intersection of artificial intelligence and human-computer interaction. This positioning offers a promising avenue to enhance the performance and adaptability of intelligent systems while also improving the alignment of their objectives with human values. The training of large language models (LLMs) has impressively demonstrated this potential in recent years, where RLHF played a decisive role in directing the model's capabilities toward human objectives. This article provides a comprehensive overview of the fundamentals of RLHF, exploring the intricate dynamics between RL agents and human input. While recent focus has been on RLHF for LLMs, our survey adopts a broader perspective, examining the diverse applications and wide-ranging impact of the technique. We delve into the core principles that underpin RLHF, shedding light on the symbiotic relationship between algorithms and human feedback, and discuss the main research trends in the field. By synthesizing the current landscape of RLHF research, this article aims to provide researchers as well as practitioners with a comprehensive understanding of this rapidly growing field of research.

Valentin Margraf, Marcel Wever, Sandra Gilhuber et al.

In settings where only a budgeted amount of labeled data can be afforded, active learning seeks to devise query strategies for selecting the most informative data points to be labeled, aiming to enhance learning algorithms' efficiency and performance. Numerous such query strategies have been proposed and compared in the active learning literature. However, the community still lacks standardized benchmarks for comparing the performance of different query strategies. This particularly holds for the combination of query strategies with different learning algorithms into active learning pipelines and examining the impact of the learning algorithm choice. To close this gap, we propose ALPBench, which facilitates the specification, execution, and performance monitoring of active learning pipelines. It has built-in measures to ensure evaluations are done reproducibly, saving exact dataset splits and hyperparameter settings of used algorithms. In total, ALPBench consists of 86 real-world tabular classification datasets and 5 active learning settings, yielding 430 active learning problems. To demonstrate its usefulness and broad compatibility with various learning algorithms and query strategies, we conduct an exemplary study evaluating 9 query strategies paired with 8 learning algorithms in 2 different settings. We provide ALPBench here: https://github.com/ValentinMargraf/ActiveLearningPipelines.

Arduin Findeis, Timo Kaufmann, Eyke Hüllermeier et al.

Feedback data is widely used for fine-tuning and evaluating state-of-the-art AI models. Pairwise text preferences, where human or AI annotators select the "better" of two options, are particularly common. Such preferences are used to train (reward) models or to rank models with aggregate statistics. For many applications it is desirable to understand annotator preferences in addition to modelling them - not least because extensive prior work has shown various unintended biases in preference datasets. Yet, preference datasets remain challenging to interpret. Neither black-box reward models nor statistics can answer why one text is preferred over another. Manual interpretation of the numerous (long) response pairs is usually equally infeasible. In this paper, we introduce the Inverse Constitutional AI (ICAI) problem, formulating the interpretation of pairwise text preference data as a compression task. In constitutional AI, a set of principles (a constitution) is used to provide feedback and fine-tune AI models. ICAI inverts this process: given a feedback dataset, we aim to extract a constitution that best enables a large language model (LLM) to reconstruct the original annotations. We propose a corresponding ICAI algorithm and validate its generated constitutions quantitatively based on annotation reconstruction accuracy on several datasets: (a) synthetic feedback data with known principles; (b) AlpacaEval cross-annotated human feedback data; (c) crowdsourced Chatbot Arena data; and (d) PRISM data from diverse demographic groups. As a short and interpretable representation of the original dataset, generated constitutions have many potential use cases: help identify undesirable annotator biases, understand model performance better, scale feedback to unseen data, or adapt models to individual user or group preferences. We release the source code at https://github.com/rdnfn/icai.

Michael Dellnitz, Eyke Hüllermeier, Marvin Lücke et al.

Many problems in science and engineering require an efficient numerical approximation of integrals or solutions to differential equations. For systems with rapidly changing dynamics, an equidistant discretization is often inadvisable as it either results in prohibitively large errors or computational effort. To this end, adaptive schemes, such as solvers based on Runge--Kutta pairs, have been developed which adapt the step size based on local error estimations at each step. While the classical schemes apply very generally and are highly efficient on regular systems, they can behave sub-optimal when an inefficient step rejection mechanism is triggered by structurally complex systems such as chaotic systems. To overcome these issues, we propose a method to tailor numerical schemes to the problem class at hand. This is achieved by combining simple, classical quadrature rules or ODE solvers with data-driven time-stepping controllers. Compared with learning solution operators to ODEs directly, it generalises better to unseen initial data as our approach employs classical numerical schemes as base methods. At the same time it can make use of identified structures of a problem class and, therefore, outperforms state-of-the-art adaptive schemes. Several examples demonstrate superior efficiency. Source code is available at https://github.com/lueckem/quadrature-ML.

Maximilian Muschalik, Fabian Fumagalli, Barbara Hammer et al.

While shallow decision trees may be interpretable, larger ensemble models like gradient-boosted trees, which often set the state of the art in machine learning problems involving tabular data, still remain black box models. As a remedy, the Shapley value (SV) is a well-known concept in explainable artificial intelligence (XAI) research for quantifying additive feature attributions of predictions. The model-specific TreeSHAP methodology solves the exponential complexity for retrieving exact SVs from tree-based models. Expanding beyond individual feature attribution, Shapley interactions reveal the impact of intricate feature interactions of any order. In this work, we present TreeSHAP-IQ, an efficient method to compute any-order additive Shapley interactions for predictions of tree-based models. TreeSHAP-IQ is supported by a mathematical framework that exploits polynomial arithmetic to compute the interaction scores in a single recursive traversal of the tree, akin to Linear TreeSHAP. We apply TreeSHAP-IQ on state-of-the-art tree ensembles and explore interactions on well-established benchmark datasets.

Mira Jürgens, Nis Meinert, Viktor Bengs et al.

Trustworthy ML systems should not only return accurate predictions, but also a reliable representation of their uncertainty. Bayesian methods are commonly used to quantify both aleatoric and epistemic uncertainty, but alternative approaches, such as evidential deep learning methods, have become popular in recent years. The latter group of methods in essence extends empirical risk minimization (ERM) for predicting second-order probability distributions over outcomes, from which measures of epistemic (and aleatoric) uncertainty can be extracted. This paper presents novel theoretical insights of evidential deep learning, highlighting the difficulties in optimizing second-order loss functions and interpreting the resulting epistemic uncertainty measures. With a systematic setup that covers a wide range of approaches for classification, regression and counts, it provides novel insights into issues of identifiability and convergence in second-order loss minimization, and the relative (rather than absolute) nature of epistemic uncertainty measures.

Moritz Herrmann, F. Julian D. Lange, Katharina Eggensperger et al.

We warn against a common but incomplete understanding of empirical research in machine learning that leads to non-replicable results, makes findings unreliable, and threatens to undermine progress in the field. To overcome this alarming situation, we call for more awareness of the plurality of ways of gaining knowledge experimentally but also of some epistemic limitations. In particular, we argue most current empirical machine learning research is fashioned as confirmatory research while it should rather be considered exploratory.

Fabian Fumagalli, Maximilian Muschalik, Patrick Kolpaczki et al.

The Shapley value (SV) is a prevalent approach of allocating credit to machine learning (ML) entities to understand black box ML models. Enriching such interpretations with higher-order interactions is inevitable for complex systems, where the Shapley Interaction Index (SII) is a direct axiomatic extension of the SV. While it is well-known that the SV yields an optimal approximation of any game via a weighted least square (WLS) objective, an extension of this result to SII has been a long-standing open problem, which even led to the proposal of an alternative index. In this work, we characterize higher-order SII as a solution to a WLS problem, which constructs an optimal approximation via SII and $k$-Shapley values ($k$-SII). We prove this representation for the SV and pairwise SII and give empirically validated conjectures for higher orders. As a result, we propose KernelSHAP-IQ, a direct extension of KernelSHAP for SII, and demonstrate state-of-the-art performance for feature interactions.

Alireza Javanmardi, David Stutz, Eyke Hüllermeier

Credal sets are sets of probability distributions that are considered as candidates for an imprecisely known ground-truth distribution. In machine learning, they have recently attracted attention as an appealing formalism for uncertainty representation, in particular due to their ability to represent both the aleatoric and epistemic uncertainty in a prediction. However, the design of methods for learning credal set predictors remains a challenging problem. In this paper, we make use of conformal prediction for this purpose. More specifically, we propose a method for predicting credal sets in the classification task, given training data labeled by probability distributions. Since our method inherits the coverage guarantees of conformal prediction, our conformal credal sets are guaranteed to be valid with high probability (without any assumptions on model or distribution). We demonstrate the applicability of our method to natural language inference, a highly ambiguous natural language task where it is common to obtain multiple annotations per example.

Paul Hofman, Yusuf Sale, Eyke Hüllermeier

Uncertainty representation and quantification are paramount in machine learning and constitute an important prerequisite for safety-critical applications. In this paper, we propose novel measures for the quantification of aleatoric and epistemic uncertainty based on proper scoring rules, which are loss functions with the meaningful property that they incentivize the learner to predict ground-truth (conditional) probabilities. We assume two common representations of (epistemic) uncertainty, namely, in terms of a credal set, i.e. a set of probability distributions, or a second-order distribution, i.e., a distribution over probability distributions. Our framework establishes a natural bridge between these representations. We provide a formal justification of our approach and introduce new measures of epistemic and aleatoric uncertainty as concrete instantiations.

Yusuf Sale, Paul Hofman, Lisa Wimmer et al.

Uncertainty quantification is a critical aspect of machine learning models, providing important insights into the reliability of predictions and aiding the decision-making process in real-world applications. This paper proposes a novel way to use variance-based measures to quantify uncertainty on the basis of second-order distributions in classification problems. A distinctive feature of the measures is the ability to reason about uncertainties on a class-based level, which is useful in situations where nuanced decision-making is required. Recalling some properties from the literature, we highlight that the variance-based measures satisfy important (axiomatic) properties. In addition to this axiomatic approach, we present empirical results showing the measures to be effective and competitive to commonly used entropy-based measures.

Julian Rodemann, Federico Croppi, Philipp Arens et al.

Bayesian optimization (BO) with Gaussian processes (GP) has become an indispensable algorithm for black box optimization problems. Not without a dash of irony, BO is often considered a black box itself, lacking ways to provide reasons as to why certain parameters are proposed to be evaluated. This is particularly relevant in human-in-the-loop applications of BO, such as in robotics. We address this issue by proposing ShapleyBO, a framework for interpreting BO's proposals by game-theoretic Shapley values.They quantify each parameter's contribution to BO's acquisition function. Exploiting the linearity of Shapley values, we are further able to identify how strongly each parameter drives BO's exploration and exploitation for additive acquisition functions like the confidence bound. We also show that ShapleyBO can disentangle the contributions to exploration into those that explore aleatoric and epistemic uncertainty. Moreover, our method gives rise to a ShapleyBO-assisted human machine interface (HMI), allowing users to interfere with BO in case proposals do not align with human reasoning. We demonstrate this HMI's benefits for the use case of personalizing wearable robotic devices (assistive back exosuits) by human-in-the-loop BO. Results suggest human-BO teams with access to ShapleyBO can achieve lower regret than teams without.

Fabian Fumagalli, Maximilian Muschalik, Eyke Hüllermeier et al.

Feature-based explanations, using perturbations or gradients, are a prevalent tool to understand decisions of black box machine learning models. Yet, differences between these methods still remain mostly unknown, which limits their applicability for practitioners. In this work, we introduce a unified framework for local and global feature-based explanations using two well-established concepts: functional ANOVA (fANOVA) from statistics, and the notion of value and interaction from cooperative game theory. We introduce three fANOVA decompositions that determine the influence of feature distributions, and use game-theoretic measures, such as the Shapley value and interactions, to specify the influence of higher-order interactions. Our framework combines these two dimensions to uncover similarities and differences between a wide range of explanation techniques for features and groups of features. We then empirically showcase the usefulness of our framework on synthetic and real-world datasets.

Christopher Bülte, Yusuf Sale, Timo Löhr et al.

Uncertainty quantification (UQ) is crucial in machine learning, yet most (axiomatic) studies of uncertainty measures focus on classification, leaving a gap in regression settings with limited formal justification and evaluations. In this work, we introduce a set of axioms to rigorously assess measures of aleatoric, epistemic, and total uncertainty in supervised regression. By utilizing a predictive exponential family, we can generalize commonly used approaches for uncertainty representation and corresponding uncertainty measures. More specifically, we analyze the widely used entropy- and variance-based measures regarding limitations and challenges. Our findings provide a principled foundation for uncertainty quantification in regression, offering theoretical insights and practical guidelines for reliable uncertainty assessment.

Michele Caprio, Yusuf Sale, Eyke Hüllermeier

Recently, Cella and Martin proved how, under an assumption called consonance, a credal set (i.e. a closed and convex set of probabilities) can be derived from the conformal transducer associated with transductive conformal prediction. We show that the Imprecise Highest Density Region (IHDR) associated with such a credal set corresponds to the classical Conformal Prediction Region. In proving this result, we establish a new relationship between Conformal Prediction and Imprecise Probability (IP) theories, via the IP concept of a cloud. A byproduct of our presentation is the discovery that consonant plausibility functions are monoid homomorphisms, a new algebraic property of an IP tool.

Kacper Sokol, Eyke Hüllermeier

This position paper argues that, to its detriment, transparency research overlooks many foundational concepts of artificial intelligence. Here, we focus on uncertainty quantification -- in the context of ante-hoc interpretability and counterfactual explainability -- showing how its adoption could address key challenges in the field. First, we posit that uncertainty and ante-hoc interpretability offer complementary views of the same underlying idea; second, we assert that uncertainty provides a principled unifying framework for counterfactual explainability. Consequently, inherently transparent models can benefit from human-centred explanatory insights -- like counterfactuals -- which are otherwise missing. At a higher level, integrating artificial intelligence fundamentals into transparency research promises to yield more reliable, robust and understandable predictive models.

Mira Jürgens, Thomas Mortier, Eyke Hüllermeier et al.

The accurate representation of epistemic uncertainty is a challenging yet essential task in machine learning. A widely used representation corresponds to convex sets of probabilistic predictors, also known as credal sets. One popular way of constructing these credal sets is via ensembling or specialized supervised learning methods, where the epistemic uncertainty can be quantified through measures such as the set size or the disagreement among members. In principle, these sets should contain the true data-generating distribution. As a necessary condition for this validity, we adopt the strongest notion of calibration as a proxy. Concretely, we propose a novel statistical test to determine whether there is a convex combination of the set's predictions that is calibrated in distribution. In contrast to previous methods, our framework allows the convex combination to be instance dependent, recognizing that different ensemble members may be better calibrated in different regions of the input space. Moreover, we learn this combination via proper scoring rules, which inherently optimize for calibration. Building on differentiable, kernel-based estimators of calibration errors, we introduce a nonparametric testing procedure and demonstrate the benefits of capturing instance-level variability on of synthetic and real-world experiments.

Maximilian Muschalik, Fabian Fumagalli, Paolo Frazzetto et al.

Albeit the ubiquitous use of Graph Neural Networks (GNNs) in machine learning (ML) prediction tasks involving graph-structured data, their interpretability remains challenging. In explainable artificial intelligence (XAI), the Shapley Value (SV) is the predominant method to quantify contributions of individual features to a ML model's output. Addressing the limitations of SVs in complex prediction models, Shapley Interactions (SIs) extend the SV to groups of features. In this work, we explain single graph predictions of GNNs with SIs that quantify node contributions and interactions among multiple nodes. By exploiting the GNN architecture, we show that the structure of interactions in node embeddings are preserved for graph prediction. As a result, the exponential complexity of SIs depends only on the receptive fields, i.e. the message-passing ranges determined by the connectivity of the graph and the number of convolutional layers. Based on our theoretical results, we introduce GraphSHAP-IQ, an efficient approach to compute any-order SIs exactly. GraphSHAP-IQ is applicable to popular message passing techniques in conjunction with a linear global pooling and output layer. We showcase that GraphSHAP-IQ substantially reduces the exponential complexity of computing exact SIs on multiple benchmark datasets. Beyond exact computation, we evaluate GraphSHAP-IQ's approximation of SIs on popular GNN architectures and compare with existing baselines. Lastly, we visualize SIs of real-world water distribution networks and molecule structures using a SI-Graph.

Mohammad Hossein Shaker, Eyke Hüllermeier

The Random Forest (RF) classifier is often claimed to be relatively well calibrated when compared with other machine learning methods. Moreover, the existing literature suggests that traditional calibration methods, such as isotonic regression, do not substantially enhance the calibration of RF probability estimates unless supplied with extensive calibration data sets, which can represent a significant obstacle in cases of limited data availability. Nevertheless, there seems to be no comprehensive study validating such claims and systematically comparing state-of-the-art calibration methods specifically for RF. To close this gap, we investigate a broad spectrum of calibration methods tailored to or at least applicable to RF, ranging from scaling techniques to more advanced algorithms. Our results based on synthetic as well as real-world data unravel the intricacies of RF probability estimates, scrutinize the impacts of hyper-parameters, compare calibration methods in a systematic way. We show that a well-optimized RF performs as well as or better than leading calibration approaches.

Leandra Fichtel, Maximilian Spliethöver, Eyke Hüllermeier et al.

The ability to generate explanations that are understood by explainees is the quintessence of explainable artificial intelligence. Since understanding depends on the explainee's background and needs, recent research focused on co-constructive explanation dialogues, where an explainer continuously monitors the explainee's understanding and adapts their explanations dynamically. We investigate the ability of large language models (LLMs) to engage as explainers in co-constructive explanation dialogues. In particular, we present a user study in which explainees interact with an LLM in two settings, one of which involves the LLM being instructed to explain a topic co-constructively. We evaluate the explainees' understanding before and after the dialogue, as well as their perception of the LLMs' co-constructive behavior. Our results suggest that LLMs show some co-constructive behaviors, such as asking verification questions, that foster the explainees' engagement and can improve understanding of a topic. However, their ability to effectively monitor the current understanding and scaffold the explanations accordingly remains limited.

Paul Hofman, Yusuf Sale, Eyke Hüllermeier

Proper quantification of predictive uncertainty is essential for the use of machine learning in safety-critical applications. Various uncertainty measures have been proposed for this purpose, typically claiming superiority over other measures. In this paper, we argue that there is no single best measure. Instead, uncertainty quantification should be tailored to the specific application. To this end, we use a flexible family of uncertainty measures that distinguishes between total, aleatoric, and epistemic uncertainty of second-order distributions. These measures can be instantiated with specific loss functions, so-called proper scoring rules, to control their characteristics, and we show that different characteristics are useful for different tasks. In particular, we show that, for the task of selective prediction, the scoring rule should ideally match the task loss. On the other hand, for out-of-distribution detection, our results confirm that mutual information, a widely used measure of epistemic uncertainty, performs best. Furthermore, in an active learning setting, epistemic uncertainty based on zero-one loss is shown to consistently outperform other uncertainty measures.

Timo Löhr, Paul Hofman, Felix Mohr et al.

Predictions in the form of sets of probability distributions, so-called credal sets, provide a suitable means to represent a learner's epistemic uncertainty. In this paper, we propose a theoretically grounded approach to credal prediction based on the statistical notion of relative likelihood: The target of prediction is the set of all (conditional) probability distributions produced by the collection of plausible models, namely those models whose relative likelihood exceeds a specified threshold. This threshold has an intuitive interpretation and allows for controlling the trade-off between correctness and precision of credal predictions. We tackle the problem of approximating credal sets defined in this way by means of suitably modified ensemble learning techniques. To validate our approach, we illustrate its effectiveness by experiments on benchmark datasets demonstrating superior uncertainty representation without compromising predictive performance. We also compare our method against several state-of-the-art baselines in credal prediction.