Matej Zečević

Matej Zečević, Moritz Willig, Devendra Singh Dhami et al.

Some argue scale is all what is needed to achieve AI, covering even causal models. We make it clear that large language models (LLMs) cannot be causal and give reason onto why sometimes we might feel otherwise. To this end, we define and exemplify a new subgroup of Structural Causal Model (SCM) that we call meta SCM which encode causal facts about other SCM within their variables. We conjecture that in the cases where LLM succeed in doing causal inference, underlying was a respective meta SCM that exposed correlations between causal facts in natural language on whose data the LLM was ultimately trained. If our hypothesis holds true, then this would imply that LLMs are like parrots in that they simply recite the causal knowledge embedded in the data. Our empirical analysis provides favoring evidence that current LLMs are even weak `causal parrots.'

Moritz Willig, Matej Zečević, Devendra Singh Dhami et al.

Foundation models are subject to an ongoing heated debate, leaving open the question of progress towards AGI and dividing the community into two camps: the ones who see the arguably impressive results as evidence to the scaling hypothesis, and the others who are worried about the lack of interpretability and reasoning capabilities. By investigating to which extent causal representations might be captured by these large scale language models, we make a humble efforts towards resolving the ongoing philosophical conflicts.

Matej Zečević, Florian Peter Busch, Devendra Singh Dhami et al.

Linear Programs (LP) are celebrated widely, particularly so in machine learning where they have allowed for effectively solving probabilistic inference tasks or imposing structure on end-to-end learning systems. Their potential might seem depleted but we propose a foundational, causal perspective that reveals intriguing intra- and inter-structure relations for LP components. We conduct a systematic, empirical investigation on general-, shortest path- and energy system LPs.

Jonas Seng, Matej Zečević, Devendra Singh Dhami et al.

Simulations are ubiquitous in machine learning. Especially in graph learning, simulations of Directed Acyclic Graphs (DAG) are being deployed for evaluating new algorithms. In the literature, it was recently argued that continuous-optimization approaches to structure discovery such as NOTEARS might be exploiting the sortability of the variable's variances in the available data due to their use of least square losses. Specifically, since structure discovery is a key problem in science and beyond, we want to be invariant to the scale being used for measuring our data (e.g. meter versus centimeter should not affect the causal direction inferred by the algorithm). In this work, we further strengthen this initial, negative empirical suggestion by both proving key results in the multivariate case and corroborating with further empirical evidence. In particular, we show that we can control the resulting graph with our targeted variance attacks, even in the case where we can only partially manipulate the variances of the data.

Salahedine Youssef, Matej Zečević, Devendra Singh Dhami et al.

Even though AI has advanced rapidly in recent years displaying success in solving highly complex problems, the class of Bongard Problems (BPs) yet remain largely unsolved by modern ML techniques. In this paper, we propose a new approach in an attempt to not only solve BPs but also extract meaning out of learned representations. This includes the reformulation of the classical BP into a reinforcement learning (RL) setting which will allow the model to gain access to counterfactuals to guide its decisions but also explain its decisions. Since learning meaningful representations in BPs is an essential sub-problem, we further make use of contrastive learning for the extraction of low level features from pixel data. Several experiments have been conducted for analyzing the general BP-RL setup, feature extraction methods and using the best combination for the feature space analysis and its interpretation.

Nils Dycke, Matej Zečević, Ilia Kuznetsov et al.

Critical text assessment is at the core of many expert activities, such as fact-checking, peer review, and essay grading. Yet, existing work treats critical text assessment as a black box problem, limiting interpretability and human-AI collaboration. To close this gap, we introduce Structured Reasoning In Critical Text Assessment (STRICTA), a novel specification framework to model text assessment as an explicit, step-wise reasoning process. STRICTA breaks down the assessment into a graph of interconnected reasoning steps drawing on causality theory (Pearl, 1995). This graph is populated based on expert interaction data and used to study the assessment process and facilitate human-AI collaboration. We formally define STRICTA and apply it in a study on biomedical paper assessment, resulting in a dataset of over 4000 reasoning steps from roughly 40 biomedical experts on more than 20 papers. We use this dataset to empirically study expert reasoning in critical text assessment, and investigate if LLMs are able to imitate and support experts within these workflows. The resulting tools and datasets pave the way for studying collaborative expert-AI reasoning in text assessment, in peer review and beyond.

Harsh Poonia, Moritz Willig, Zhongjie Yu et al.

Causal inference in hybrid domains, characterized by a mixture of discrete and continuous variables, presents a formidable challenge. We take a step towards this direction and propose Characteristic Interventional Sum-Product Network ($χ$SPN) that is capable of estimating interventional distributions in presence of random variables drawn from mixed distributions. $χ$SPN uses characteristic functions in the leaves of an interventional SPN (iSPN) thereby providing a unified view for discrete and continuous random variables through the Fourier-Stieltjes transform of the probability measures. A neural network is used to estimate the parameters of the learned iSPN using the intervened data. Our experiments on 3 synthetic heterogeneous datasets suggest that $χ$SPN can effectively capture the interventional distributions for both discrete and continuous variables while being expressive and causally adequate. We also show that $χ$SPN generalize to multiple interventions while being trained only on a single intervention data.

Matej Zečević, Moritz Willig, Devendra Singh Dhami et al.

Many researchers have voiced their support towards Pearl's counterfactual theory of causation as a stepping stone for AI/ML research's ultimate goal of intelligent systems. As in any other growing subfield, patience seems to be a virtue since significant progress on integrating notions from both fields takes time, yet, major challenges such as the lack of ground truth benchmarks or a unified perspective on classical problems such as computer vision seem to hinder the momentum of the research movement. This present work exemplifies how the Pearl Causal Hierarchy (PCH) can be understood on image data by providing insights on several intricacies but also challenges that naturally arise when applying key concepts from Pearlian causality to the study of image data.

Florian Peter Busch, Matej Zečević, Kristian Kersting et al.

Linear Programs (LPs) have been one of the building blocks in machine learning and have championed recent strides in differentiable optimizers for learning systems. While there exist solvers for even high-dimensional LPs, understanding said high-dimensional solutions poses an orthogonal and unresolved problem. We introduce an approach where we consider neural encodings for LPs that justify the application of attribution methods from explainable artificial intelligence (XAI) designed for neural learning systems. The several encoding functions we propose take into account aspects such as feasibility of the decision space, the cost attached to each input, or the distance to special points of interest. We investigate the mathematical consequences of several XAI methods on said neural LP encodings. We empirically show that the attribution methods Saliency and LIME reveal indistinguishable results up to perturbation levels, and we propose the property of Directedness as the main discriminative criterion between Saliency and LIME on one hand, and a perturbation-based Feature Permutation approach on the other hand. Directedness indicates whether an attribution method gives feature attributions with respect to an increase of that feature. We further notice the baseline selection problem beyond the classical computer vision setting for Integrated Gradients.

David Steinmann, Matej Zečević, Devendra Singh Dhami et al.

There has been a recent push in making machine learning models more interpretable so that their performance can be trusted. Although successful, these methods have mostly focused on the deep learning methods while the fundamental optimization methods in machine learning such as linear programs (LP) have been left out. Even if LPs can be considered as whitebox or clearbox models, they are not easy to understand in terms of relationships between inputs and outputs. As a linear program only provides the optimal solution to an optimization problem, further explanations are often helpful. In this work, we extend the attribution methods for explaining neural networks to linear programs. These methods explain the model by providing relevance scores for the model inputs, to show the influence of each input on the output. Alongside using classical gradient-based attribution methods we also propose a way to adapt perturbation-based attribution methods to LPs. Our evaluations of several different linear and integer problems showed that attribution methods can generate useful explanations for linear programs. However, we also demonstrate that using a neural attribution method directly might come with some drawbacks, as the properties of these methods on neural networks do not necessarily transfer to linear programs. The methods can also struggle if a linear program has more than one optimal solution, as a solver just returns one possible solution. Our results can hopefully be used as a good starting point for further research in this direction.

Moritz Willig, Matej Zečević, Devendra Singh Dhami et al.

Structural causal models (SCMs) are a powerful tool for understanding the complex causal relationships that underlie many real-world systems. As these systems grow in size, the number of variables and complexity of interactions between them does, too. Thus, becoming convoluted and difficult to analyze. This is particularly true in the context of machine learning and artificial intelligence, where an ever increasing amount of data demands for new methods to simplify and compress large scale SCM. While methods for marginalizing and abstracting SCM already exist today, they may destroy the causality of the marginalized model. To alleviate this, we introduce the concept of consolidating causal mechanisms to transform large-scale SCM while preserving consistent interventional behaviour. We show consolidation is a powerful method for simplifying SCM, discuss reduction of computational complexity and give a perspective on generalizing abilities of consolidated SCM.

Kieran Didi, Matej Zečević

Research around AI for Science has seen significant success since the rise of deep learning models over the past decade, even with longstanding challenges such as protein structure prediction. However, this fast development inevitably made their flaws apparent -- especially in domains of reasoning where understanding the cause-effect relationship is important. One such domain is drug discovery, in which such understanding is required to make sense of data otherwise plagued by spurious correlations. Said spuriousness only becomes worse with the ongoing trend of ever-increasing amounts of data in the life sciences and thereby restricts researchers in their ability to understand disease biology and create better therapeutics. Therefore, to advance the science of drug discovery with AI it is becoming necessary to formulate the key problems in the language of causality, which allows the explication of modelling assumptions needed for identifying true cause-effect relationships. In this attention paper, we present causal drug discovery as the craft of creating models that ground the process of drug discovery in causal reasoning.

Matej Zečević, Moritz Willig, Jonas Seng et al.

This short paper discusses continually updated causal abstractions as a potential direction of future research. The key idea is to revise the existing level of causal abstraction to a different level of detail that is both consistent with the history of observed data and more effective in solving a given task.

Sebastian Rödling, Matej Zečević, Devendra Singh Dhami et al.

Structural Causal Explanations (SCEs) can be used to automatically generate explanations in natural language to questions about given data that are grounded in a (possibly learned) causal model. Unfortunately they work for small data only. In turn they are not attractive to offer reasons for events, e.g., tracking causal changes over multiple time steps, or a behavioral component that involves feedback loops through actions of an agent. To this end, we generalize SCEs to a (recursive) formulation of explanation trees to capture the temporal interactions between reasons. We show the benefits of this more general SCE algorithm on synthetic time-series data and a 2D grid game, and further compare it to the base SCE and other existing methods for causal explanations.

Moritz Willig, Matej Zečević, Devendra Singh Dhami et al.

Most algorithms in classical and contemporary machine learning focus on correlation-based dependence between features to drive performance. Although success has been observed in many relevant problems, these algorithms fail when the underlying causality is inconsistent with the assumed relations. We propose a novel model-agnostic loss function called Causal Loss that improves the interventional quality of the prediction using an intervened neural-causal regularizer. In support of our theoretical results, our experimental illustration shows how causal loss bestows a non-causal associative model (like a standard neural net or decision tree) with interventional capabilities.

Matej Zečević, Devendra Singh Dhami, Kristian Kersting

Neurally-parameterized Structural Causal Models in the Pearlian notion to causality, referred to as NCM, were recently introduced as a step towards next-generation learning systems. However, said NCM are only concerned with the learning aspect of causal inference but totally miss out on the architecture aspect. That is, actual causal inference within NCM is intractable in that the NCM won't return an answer to a query in polynomial time. This insight follows as corollary to the more general statement on the intractability of arbitrary SCM parameterizations, which we prove in this work through classical 3-SAT reduction. Since future learning algorithms will be required to deal with both high dimensional data and highly complex mechanisms governing the data, we ultimately believe work on tractable inference for causality to be decisive. We also show that not all ``causal'' models are created equal. More specifically, there are models capable of answering causal queries that are not SCM, which we refer to as \emph{partially causal models} (PCM). We provide a tabular taxonomy in terms of tractability properties for all of the different model families, namely correlation-based, PCM and SCM. To conclude our work, we also provide some initial ideas on how to overcome parts of the intractability of causal inference with SCM by showing an example of how parameterizing an SCM with SPN modules can at least allow for tractable mechanisms. We hope that our impossibility result alongside the taxonomy for tractability in causal models can raise awareness for this novel research direction since achieving success with causality in real world downstream tasks will not only depend on learning correct models as we also require having the practical ability to gain access to model inferences.

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.

Matej Zečević, Devendra Singh Dhami, Petar Veličković et al.

Causality can be described in terms of a structural causal model (SCM) that carries information on the variables of interest and their mechanistic relations. For most processes of interest the underlying SCM will only be partially observable, thus causal inference tries leveraging the exposed. Graph neural networks (GNN) as universal approximators on structured input pose a viable candidate for causal learning, suggesting a tighter integration with SCM. To this effect we present a theoretical analysis from first principles that establishes a more general view on neural-causal models, revealing several novel connections between GNN and SCM. We establish a new model class for GNN-based causal inference that is necessary and sufficient for causal effect identification. Our empirical illustration on simulations and standard benchmarks validate our theoretical proofs.

Matej Zečević, Devendra Singh Dhami, Kristian Kersting

The recent years have been marked by extended research on adversarial attacks, especially on deep neural networks. With this work we intend on posing and investigating the question of whether the phenomenon might be more general in nature, that is, adversarial-style attacks outside classical classification tasks. Specifically, we investigate optimization problems as they constitute a fundamental part of modern AI research. To this end, we consider the base class of optimizers namely Linear Programs (LPs). On our initial attempt of a naïve mapping between the formalism of adversarial examples and LPs, we quickly identify the key ingredients missing for making sense of a reasonable notion of adversarial examples for LPs. Intriguingly, the formalism of Pearl's notion to causality allows for the right description of adversarial like examples for LPs. Characteristically, we show the direct influence of the Structural Causal Model (SCM) onto the subsequent LP optimization, which ultimately exposes a notion of confounding in LPs (inherited by said SCM) that allows for adversarial-style attacks. We provide both the general proof formally alongside existential proofs of such intriguing LP-parameterizations based on SCM for three combinatorial problems, namely Linear Assignment, Shortest Path and a real world problem of energy systems.

Matej Zečević, Devendra Singh Dhami, Athresh Karanam et al.

While probabilistic models are an important tool for studying causality, doing so suffers from the intractability of inference. As a step towards tractable causal models, we consider the problem of learning interventional distributions using sum-product networks (SPNs) that are over-parameterized by gate functions, e.g., neural networks. Providing an arbitrarily intervened causal graph as input, effectively subsuming Pearl's do-operator, the gate function predicts the parameters of the SPN. The resulting interventional SPNs are motivated and illustrated by a structural causal model themed around personal health. Our empirical evaluation on three benchmark data sets as well as a synthetic health data set clearly demonstrates that interventional SPNs indeed are both expressive in modelling and flexible in adapting to the interventions.