David Jensen

Pracheta Amaranath, Anant Bhide, David Jensen et al.

Metamodels for discrete-event simulations approximate the behavior of simulation models without running expensive simulations. Prior work introduced modular dynamic Bayesian networks (MDBNs) -- a class of metamodels that can estimate a range of probabilistic and causal queries (PCQs) using a single, trained model -- but the method was limited to Markovian systems. In this paper, we initiate an extension of MDBNs to non-Markovian queues by approximating non-exponential distributions using phase-type distributions. This approach raises novel challenges, including balancing metamodeling accuracy and tractability when choosing the number of phases, efficiently learning metamodel parameters, and choosing the sampling interval that is used to approximate a continuous-time simulation by a discrete-time MDBN. We provide preliminary solutions to these challenges, yielding the first causal metamodeling technique for non-Markovian systems. Experiments on a G/M/1 queue demonstrate that the MDBN can produce accurate answers to PCQs with orders-of-magnitude speedup of inference times relative to direct simulation.

Erica Cai, Andrew McGregor, David Jensen

Learning causal structure is useful in many areas of artificial intelligence, including planning, robotics, and explanation. Constraint-based structure learning algorithms such as PC use conditional independence (CI) tests to infer causal structure. Traditionally, constraint-based algorithms perform CI tests with a preference for smaller-sized conditioning sets, partially because the statistical power of conventional CI tests declines rapidly as the size of the conditioning set increases. However, many modern conditional independence tests are model-based, and these tests use well-regularized models that maintain statistical power even with very large conditioning sets. This suggests an intriguing new strategy for constraint-based algorithms which may result in a reduction of the total number of CI tests performed: Test variable pairs with large conditioning sets first, as a pre-processing step that finds some conditional independencies quickly, before moving on to the more conventional strategy that favors small conditioning sets. We propose such a pre-processing step for the PC algorithm which relies on performing CI tests on a few randomly selected large conditioning sets. We perform an empirical analysis on directed acyclic graphs (DAGs) that correspond to real-world systems and both empirical and theoretical analyses for Erdős-Renyi DAGs. Our results show that Pre-Processing Plus PC (P3PC) performs far fewer CI tests than the original PC algorithm, between 0.5% to 36%, and often less than 10%, of the CI tests that the PC algorithm alone performs. The efficiency gains are particularly significant for the DAGs corresponding to real-world systems.

David Jensen, Brian LaMacchia, Ufuk Topcu et al.

Algorithmic robustness refers to the sustained performance of a computational system in the face of change in the nature of the environment in which that system operates or in the task that the system is meant to perform. Below, we motivate the importance of algorithmic robustness, present a conceptual framework, and highlight the relevant areas of research for which algorithmic robustness is relevant. Why robustness? Robustness is an important enabler of other goals that are frequently cited in the context of public policy decisions about computational systems, including trustworthiness, accountability, fairness, and safety. Despite this dependence, it tends to be under-recognized compared to these other concepts. This is unfortunate, because robustness is often more immediately achievable than these other ultimate goals, which can be more subjective and exacting. Thus, we highlight robustness as an important goal for researchers, engineers, regulators, and policymakers when considering the design, implementation, and deployment of computational systems. We urge researchers and practitioners to elevate the attention paid to robustness when designing and evaluating computational systems. For many key systems, the immediate question after any demonstration of high performance should be: "How robust is that performance to realistic changes in the task or environment?" Greater robustness will set the stage for systems that are more trustworthy, accountable, fair, and safe. Toward that end, this document provides a brief roadmap to some of the concepts and existing research around the idea of algorithmic robustness.

Katherine Avery, Jack Kenney, Pracheta Amaranath et al.

Recent work in reinforcement learning has focused on several characteristics of learned policies that go beyond maximizing reward. These properties include fairness, explainability, generalization, and robustness. In this paper, we define interventional robustness (IR), a measure of how much variability is introduced into learned policies by incidental aspects of the training procedure, such as the order of training data or the particular exploratory actions taken by agents. A training procedure has high IR when the agents it produces take very similar actions under intervention, despite variation in these incidental aspects of the training procedure. We develop an intuitive, quantitative measure of IR and calculate it for eight algorithms in three Atari environments across dozens of interventions and states. From these experiments, we find that IR varies with the amount of training and type of algorithm and that high performance does not imply high IR, as one might expect.

Emma Tosch, Kaleigh Clary, John Foley et al.

Evaluation of deep reinforcement learning (RL) is inherently challenging. In particular, learned policies are largely opaque, and hypotheses about the behavior of deep RL agents are difficult to test in black-box environments. Considerable effort has gone into addressing opacity, but almost no effort has been devoted to producing high quality environments for experimental evaluation of agent behavior. We present TOYBOX, a new high-performance, open-source* subset of Atari environments re-designed for the experimental evaluation of deep RL. We show that TOYBOX enables a wide range of experiments and analyses that are impossible in other environments. *https://kdl-umass.github.io/Toybox/

Jatin Nainani, Sankaran Vaidyanathan, AJ Yeung et al.

Mechanistic interpretability aims to understand the inner workings of large neural networks by identifying circuits, or minimal subgraphs within the model that implement algorithms responsible for performing specific tasks. These circuits are typically discovered and analyzed using a narrowly defined prompt format. However, given the abilities of large language models (LLMs) to generalize across various prompt formats for the same task, it remains unclear how well these circuits generalize. For instance, it is unclear whether the models generalization results from reusing the same circuit components, the components behaving differently, or the use of entirely different components. In this paper, we investigate the generality of the indirect object identification (IOI) circuit in GPT-2 small, which is well-studied and believed to implement a simple, interpretable algorithm. We evaluate its performance on prompt variants that challenge the assumptions of this algorithm. Our findings reveal that the circuit generalizes surprisingly well, reusing all of its components and mechanisms while only adding additional input edges. Notably, the circuit generalizes even to prompt variants where the original algorithm should fail; we discover a mechanism that explains this which we term S2 Hacking. Our findings indicate that circuits within LLMs may be more flexible and general than previously recognized, underscoring the importance of studying circuit generalization to better understand the broader capabilities of these models.

Caleb Chuck, Sankaran Vaidyanathan, Stephen Giguere et al.

Reinforcement learning (RL) algorithms often struggle to learn policies that generalize to novel situations due to issues such as causal confusion, overfitting to irrelevant factors, and failure to isolate control of state factors. These issues stem from a common source: a failure to accurately identify and exploit state-specific causal relationships in the environment. While some prior works in RL aim to identify these relationships explicitly, they rely on informal domain-specific heuristics such as spatial and temporal proximity. Actual causality offers a principled and general framework for determining the causes of particular events. However, existing definitions of actual cause often attribute causality to a large number of events, even if many of them rarely influence the outcome. Prior work on actual causality proposes normality as a solution to this problem, but its existing implementations are challenging to scale to complex and continuous-valued RL environments. This paper introduces functional actual cause (FAC), a framework that uses context-specific independencies in the environment to restrict the set of actual causes. We additionally introduce Joint Optimization for Actual Cause Inference (JACI), an algorithm that learns from observational data to infer functional actual causes. We demonstrate empirically that FAC agrees with known results on a suite of examples from the actual causality literature, and JACI identifies actual causes with significantly higher accuracy than existing heuristic methods in a set of complex, continuous-valued environments.

Isha Karn, David Jensen

Graph neural networks (GNNs) are widely used for the detection of fake news by modeling the content and propagation structure of news articles on social media. We show that two of the most commonly used benchmark data sets - GossipCop and PolitiFact - are poorly suited to evaluating the utility of models that use propagation structure. Specifically, these data sets exhibit shallow, ego-like graph topologies that provide little or no ability to differentiate among modeling methods. We systematically benchmark five GNN architectures against a structure-agnostic multilayer perceptron (MLP) that uses the same node features. We show that MLPs match or closely trail the performance of GNNs, with performance gaps often within 1-2% and overlapping confidence intervals. To isolate the contribution of structure in these datasets, we conduct controlled experiments where node features are shuffled or edge structures randomized. We find that performance collapses under feature shuffling but remains stable under edge randomization. This suggests that structure plays a negligible role in these benchmarks. Structural analysis further reveals that over 75% of nodes are only one hop from the root, exhibiting minimal structural diversity. In contrast, on synthetic datasets where node features are noisy and structure is informative, GNNs significantly outperform MLPs. These findings provide strong evidence that widely used benchmarks do not meaningfully test the utility of modeling structural features, and they motivate the development of datasets with richer, more diverse graph topologies.

Trang Nguyen, Jackson Michaels, Madalina Fiterau et al.

This paper highlights the challenge of decomposing conflict detection from conflict resolution in Vision-Language Models (VLMs) and presents potential approaches, including using a supervised metric via linear probes and group-based attention pattern analysis. We conduct a mechanistic investigation of LLaVA-OV-7B, a state-of-the-art VLM that exhibits diverse resolution behaviors when faced with conflicting multimodal inputs. Our results show that a linearly decodable conflict signal emerges in the model's intermediate layers and that attention patterns associated with conflict detection and resolution diverge at different stages of the network. These findings support the hypothesis that detection and resolution are functionally distinct mechanisms. We discuss how such decomposition enables more actionable interpretability and targeted interventions for improving model robustness in challenging multimodal settings.

Katherine Avery, Chinmay Pendse, David Jensen

Distributionally robust evaluation estimates the worst-case expected return over an uncertainty set of possible covariate and reward distributions, and distributionally robust learning finds a policy that maximizes that worst-case return across that uncertainty set. Unfortunately, current methods for distributionally robust evaluation and learning create overly conservative evaluations and policies. In this work, we propose a practical bandit evaluation and learning algorithm that tailors the uncertainty set to specific problems using mathematical programs constrained by structural equation models. Further, we show how conditional independence testing can be used to detect shifted variables for modeling. We find that the structural equation model (SEM) approach gives more accurate evaluations and learns lower-variance policies than traditional approaches, particularly for large shifts. Further, the SEM approach learns an optimal policy, assuming the model is sufficiently well-specified.

Purva Pruthi, David Jensen

Many real-world systems can be usefully represented as sets of interacting components. Examples include computational systems, such as query processors and compilers, natural systems, such as cells and ecosystems, and social systems, such as families and organizations. However, current approaches to estimating potential outcomes and causal effects typically treat such systems as single units, represent them with a fixed set of variables, and assume a homogeneous data-generating process. In this work, we study a compositional approach for estimating individual-level potential outcomes and causal effects in structured systems, where each unit is represented by an instance-specific composition of multiple heterogeneous components. The compositional approach decomposes unit-level causal queries into more fine-grained queries, explicitly modeling how unit-level interventions affect component-level outcomes to generate a unit's outcome. We demonstrate this approach using modular neural network architectures and show that it provides benefits for causal effect estimation from observational data, such as accurate causal effect estimation for structured units, increased sample efficiency, improved overlap between treatment and control groups, and compositional generalization to units with unseen combinations of components. Remarkably, our results show that compositional modeling can improve the accuracy of causal estimation even when component-level outcomes are unobserved. We also create and use a set of real-world evaluation environments for the empirical evaluation of compositional approaches for causal effect estimation and demonstrate the role of composition structure, varying amounts of component-level data access, and component heterogeneity in the performance of compositional models as compared to the non-compositional approaches.

Jeff Druce, James Niehaus, Vanessa Moody et al.

The advances in artificial intelligence enabled by deep learning architectures are undeniable. In several cases, deep neural network driven models have surpassed human level performance in benchmark autonomy tasks. The underlying policies for these agents, however, are not easily interpretable. In fact, given their underlying deep models, it is impossible to directly understand the mapping from observations to actions for any reasonably complex agent. Producing this supporting technology to "open the black box" of these AI systems, while not sacrificing performance, was the fundamental goal of the DARPA XAI program. In our journey through this program, we have several "big picture" takeaways: 1) Explanations need to be highly tailored to their scenario; 2) many seemingly high performing RL agents are extremely brittle and are not amendable to explanation; 3) causal models allow for rich explanations, but how to present them isn't always straightforward; and 4) human subjects conjure fantastically wrong mental models for AIs, and these models are often hard to break. This paper discusses the origins of these takeaways, provides amplifying information, and suggestions for future work.

Sam Witty, David Jensen, Vikash Mansinghka

A growing family of approaches to causal inference rely on Bayesian formulations of assumptions that go beyond causal graph structure. For example, Bayesian approaches have been developed for analyzing instrumental variable designs, regression discontinuity designs, and within-subjects designs. This paper introduces simulation-based identifiability (SBI), a procedure for testing the identifiability of queries in Bayesian causal inference approaches that are implemented as probabilistic programs. SBI complements analytical approaches to identifiability, leveraging a particle-based optimization scheme on simulated data to determine identifiability for analytically intractable models. We analyze SBI's soundness for a broad class of differentiable, finite-dimensional probabilistic programs with bounded effects. Finally, we provide an implementation of SBI using stochastic gradient descent, and show empirically that it agrees with known identification results on a suite of graph-based and quasi-experimental design benchmarks, including those using Gaussian processes.

Akanksha Atrey, Prashant Shenoy, David Jensen

The ubiquity of mobile devices has led to the proliferation of mobile services that provide personalized and context-aware content to their users. Modern mobile services are distributed between end-devices, such as smartphones, and remote servers that reside in the cloud. Such services thrive on their ability to predict future contexts to pre-fetch content or make context-specific recommendations. An increasingly common method to predict future contexts, such as location, is via machine learning (ML) models. Recent work in context prediction has focused on ML model personalization where a personalized model is learned for each individual user in order to tailor predictions or recommendations to a user's mobile behavior. While the use of personalized models increases efficacy of the mobile service, we argue that it increases privacy risk since a personalized model encodes contextual behavior unique to each user. To demonstrate these privacy risks, we present several attribute inference-based privacy attacks and show that such attacks can leak privacy with up to 78% efficacy for top-3 predictions. We present Pelican, a privacy-preserving personalization system for context-aware mobile services that leverages both device and cloud resources to personalize ML models while minimizing the risk of privacy leakage for users. We evaluate Pelican using real world traces for location-aware mobile services and show that Pelican can substantially reduce privacy leakage by up to 75%.

Amanda Gentzel, Purva Pruthi, David Jensen

Methods that infer causal dependence from observational data are central to many areas of science, including medicine, economics, and the social sciences. A variety of theoretical properties of these methods have been proven, but empirical evaluation remains a challenge, largely due to the lack of observational data sets for which treatment effect is known. We describe and analyze observational sampling from randomized controlled trials (OSRCT), a method for evaluating causal inference methods using data from randomized controlled trials (RCTs). This method can be used to create constructed observational data sets with corresponding unbiased estimates of treatment effect, substantially increasing the number of data sets available for empirical evaluation of causal inference methods. We show that, in expectation, OSRCT creates data sets that are equivalent to those produced by randomly sampling from empirical data sets in which all potential outcomes are available. We then perform a large-scale evaluation of seven causal inference methods over 37 data sets, drawn from RCTs, as well as simulators, real-world computational systems, and observational data sets augmented with a synthetic response variable. We find notable performance differences when comparing across data from different sources, demonstrating the importance of using data from a variety of sources when evaluating any causal inference method.

Sam Witty, Kenta Takatsu, David Jensen et al.

Latent confounders---unobserved variables that influence both treatment and outcome---can bias estimates of causal effects. In some cases, these confounders are shared across observations, e.g. all students taking a course are influenced by the course's difficulty in addition to any educational interventions they receive individually. This paper shows how to semiparametrically model latent confounders that have this structure and thereby improve estimates of causal effects. The key innovations are a hierarchical Bayesian model, Gaussian processes with structured latent confounders (GP-SLC), and a Monte Carlo inference algorithm for this model based on elliptical slice sampling. GP-SLC provides principled Bayesian uncertainty estimates of individual treatment effect with minimal assumptions about the functional forms relating confounders, covariates, treatment, and outcome. Finally, this paper shows GP-SLC is competitive with or more accurate than widely used causal inference techniques on three benchmark datasets, including the Infant Health and Development Program and a dataset showing the effect of changing temperatures on state-wide energy consumption across New England.

Katherine A. Keith, David Jensen, Brendan O'Connor

Many applications of computational social science aim to infer causal conclusions from non-experimental data. Such observational data often contains confounders, variables that influence both potential causes and potential effects. Unmeasured or latent confounders can bias causal estimates, and this has motivated interest in measuring potential confounders from observed text. For example, an individual's entire history of social media posts or the content of a news article could provide a rich measurement of multiple confounders. Yet, methods and applications for this problem are scattered across different communities and evaluation practices are inconsistent. This review is the first to gather and categorize these examples and provide a guide to data-processing and evaluation decisions. Despite increased attention on adjusting for confounding using text, there are still many open problems, which we highlight in this paper.

Akanksha Atrey, Kaleigh Clary, David Jensen

Saliency maps are frequently used to support explanations of the behavior of deep reinforcement learning (RL) agents. However, a review of how saliency maps are used in practice indicates that the derived explanations are often unfalsifiable and can be highly subjective. We introduce an empirical approach grounded in counterfactual reasoning to test the hypotheses generated from saliency maps and assess the degree to which they correspond to the semantics of RL environments. We use Atari games, a common benchmark for deep RL, to evaluate three types of saliency maps. Our results show the extent to which existing claims about Atari games can be evaluated and suggest that saliency maps are best viewed as an exploratory tool rather than an explanatory tool.

Sam Witty, Alexander Lew, David Jensen et al.

Causal inference can be formalized as Bayesian inference that combines a prior distribution over causal models and likelihoods that account for both observations and interventions. We show that it is possible to implement this approach using a sufficiently expressive probabilistic programming language. Priors are represented using probabilistic programs that generate source code in a domain specific language. Interventions are represented using probabilistic programs that edit this source code to modify the original generative process. This approach makes it straightforward to incorporate data from atomic interventions, as well as shift interventions, variance-scaling interventions, and other interventions that modify causal structure. This approach also enables the use of general-purpose inference machinery for probabilistic programs to infer probable causal structures and parameters from data. This abstract describes a prototype of this approach in the Gen probabilistic programming language.

Amanda Gentzel, Dan Garant, David Jensen

Causal inference is central to many areas of artificial intelligence, including complex reasoning, planning, knowledge-base construction, robotics, explanation, and fairness. An active community of researchers develops and enhances algorithms that learn causal models from data, and this work has produced a series of impressive technical advances. However, evaluation techniques for causal modeling algorithms have remained somewhat primitive, limiting what we can learn from experimental studies of algorithm performance, constraining the types of algorithms and model representations that researchers consider, and creating a gap between theory and practice. We argue for more frequent use of evaluation techniques that examine interventional measures rather than structural or observational measures, and that evaluate those measures on empirical data rather than synthetic data. We survey the current practice in evaluation and show that the techniques we recommend are rarely used in practice. We show that such techniques are feasible and that data sets are available to conduct such evaluations. We also show that these techniques produce substantially different results than using structural measures and synthetic data.

Kaleigh Clary, Emma Tosch, John Foley et al.

Reproducibility in reinforcement learning is challenging: uncontrolled stochasticity from many sources, such as the learning algorithm, the learned policy, and the environment itself have led researchers to report the performance of learned agents using aggregate metrics of performance over multiple random seeds for a single environment. Unfortunately, there are still pernicious sources of variability in reinforcement learning agents that make reporting common summary statistics an unsound metric for performance. Our experiments demonstrate the variability of common agents used in the popular OpenAI Baselines repository. We make the case for reporting post-training agent performance as a distribution, rather than a point estimate.

Sam Witty, Jun Ki Lee, Emma Tosch et al.

Deep reinforcement-learning methods have achieved remarkable performance on challenging control tasks. Observations of the resulting behavior give the impression that the agent has constructed a generalized representation that supports insightful action decisions. We re-examine what is meant by generalization in RL, and propose several definitions based on an agent's performance in on-policy, off-policy, and unreachable states. We propose a set of practical methods for evaluating agents with these definitions of generalization. We demonstrate these techniques on a common benchmark task for deep RL, and we show that the learned networks make poor decisions for states that differ only slightly from on-policy states, even though those states are not selected adversarially. Taken together, these results call into question the extent to which deep Q-networks learn generalized representations, and suggest that more experimentation and analysis is necessary before claims of representation learning can be supported.

John Foley, Emma Tosch, Kaleigh Clary et al.

It is a widely accepted principle that software without tests has bugs. Testing reinforcement learning agents is especially difficult because of the stochastic nature of both agents and environments, the complexity of state-of-the-art models, and the sequential nature of their predictions. Recently, the Arcade Learning Environment (ALE) has become one of the most widely used benchmark suites for deep learning research, and state-of-the-art Reinforcement Learning (RL) agents have been shown to routinely equal or exceed human performance on many ALE tasks. Since ALE is based on emulation of original Atari games, the environment does not provide semantically meaningful representations of internal game state. This means that ALE has limited utility as an environment for supporting testing or model introspection. We propose ToyBox, a collection of reimplementations of these games that solves this critical problem and enables robust testing of RL agents.

Dan Garant, David Jensen

The predominant method for evaluating the quality of causal models is to measure the graphical accuracy of the learned model structure. We present an alternative method for evaluating causal models that directly measures the accuracy of estimated interventional distributions. We contrast such distributional measures with structural measures, such as structural Hamming distance and structural intervention distance, showing that structural measures often correspond poorly to the accuracy of estimated interventional distributions. We use a number of real and synthetic datasets to illustrate various scenarios in which structural measures provide misleading results with respect to algorithm selection and parameter tuning, and we recommend that distributional measures become the new standard for evaluating causal models.

Katerina Marazopoulou, Rumi Ghosh, Prasanth Lade et al.

Yield and quality improvement is of paramount importance to any manufacturing company. One of the ways of improving yield is through discovery of the root causal factors affecting yield. We propose the use of data-driven interpretable causal models to identify key factors affecting yield. We focus on factors that are measured in different stages of production and testing in the manufacturing cycle of a product. We apply causal structure learning techniques on real data collected from this line. Specifically, the goal of this work is to learn interpretable causal models from observational data produced by manufacturing lines. Emphasis has been given to the interpretability of the models to make them actionable in the field of manufacturing. We highlight the challenges presented by assembly line data and propose ways to alleviate them.We also identify unique characteristics of data originating from assembly lines and how to leverage them in order to improve causal discovery. Standard evaluation techniques for causal structure learning shows that the learned causal models seem to closely represent the underlying latent causal relationship between different factors in the production process. These results were also validated by manufacturing domain experts who found them promising. This work demonstrates how data mining and knowledge discovery can be used for root cause analysis in the domain of manufacturing and connected industry.

James Atwood, Don Towsley, Krista Gile et al.

We investigate the problem of learning to generate complex networks from data. Specifically, we consider whether deep belief networks, dependency networks, and members of the exponential random graph family can learn to generate networks whose complex behavior is consistent with a set of input examples. We find that the deep model is able to capture the complex behavior of small networks, but that no model is able capture this behavior for networks with more than a handful of nodes.

Kun Tu, Bruno Ribeiro, Hua Jiang et al.

Recommendation systems for online dating have recently attracted much attention from the research community. In this paper we proposed a two-side matching framework for online dating recommendations and design an LDA model to learn the user preferences from the observed user messaging behavior and user profile features. Experimental results using data from a large online dating website shows that two-sided matching improves significantly the rate of successful matches by as much as 45%. Finally, using simulated matchings we show that the the LDA model can correctly capture user preferences.

Marc Maier, Katerina Marazopoulou, David Arbour et al.

The PC algorithm learns maximally oriented causal Bayesian networks. However, there is no equivalent complete algorithm for learning the structure of relational models, a more expressive generalization of Bayesian networks. Recent developments in the theory and representation of relational models support lifted reasoning about conditional independence. This enables a powerful constraint for orienting bivariate dependencies and forms the basis of a new algorithm for learning structure. We present the relational causal discovery (RCD) algorithm that learns causal relational models. We prove that RCD is sound and complete, and we present empirical results that demonstrate effectiveness.

Marc Maier, Katerina Marazopoulou, David Jensen

We extend the theory of d-separation to cases in which data instances are not independent and identically distributed. We show that applying the rules of d-separation directly to the structure of probabilistic models of relational data inaccurately infers conditional independence. We introduce relational d-separation, a theory for deriving conditional independence facts from relational models. We provide a new representation, the abstract ground graph, that enables a sound, complete, and computationally efficient method for answering d-separation queries about relational models, and we present empirical results that demonstrate effectiveness.

Marc Maier, David Jensen

The rules of d-separation provide a framework for deriving conditional independence facts from model structure. However, this theory only applies to simple directed graphical models. We introduce relational d-separation, a theory for deriving conditional independence in relational models. We provide a sound, complete, and computationally efficient method for relational d-separation, and we present empirical results that demonstrate effectiveness.