David Jensen

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.

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.

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.

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.

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%.

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.

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.

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.