Lucas Monteiro Paes, Dennis Wei, Hyo Jin Do, Hendrik Strobelt, Ronny Luss, Amit Dhurandhar, Manish Nagireddy, Karthikeyan Natesan Ramamurthy, Prasanna Sattigeri, Werner Geyer, Soumya Ghosh. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Training data attribution (TDA) is concerned with understanding model behavior in terms of the training data. This paper draws attention to the common setting where one has access only to the final trained model, and not the training algorithm or intermediate information from training. We reframe the problem in this "final-model-only" setting as one of measuring sensitivity of the model to training instances. To operationalize this reframing, we propose *further training*, with appropriate adjustment and averaging, as a gold standard method to measure sensitivity. We then unify existing gradient-based methods for TDA by showing that they all approximate the further training gold standard in different ways. We investigate empirically the quality of these gradient-based approximations to further training, for tabular, image, and text datasets and models. We find that the approximation quality of first-order methods is sometimes high but decays with the amount of further training. In contrast, the approximations given by influence function methods are more stable but surprisingly lower in quality.

We consider the issue of calibration in large language models (LLM). Recent studies have found that common interventions such as instruction tuning often result in poorly calibrated LLMs. Although calibration is well-explored in traditional applications, calibrating LLMs is uniquely challenging. These challenges stem as much from the severe computational requirements of LLMs as from their versatility, which allows them to be applied to diverse tasks. Addressing these challenges, we propose THERMOMETER, a calibration approach tailored to LLMs. THERMOMETER learns an auxiliary model, given data from multiple tasks, for calibrating a LLM. It is computationally efficient, preserves the accuracy of the LLM, and produces better-calibrated responses for new tasks. Extensive empirical evaluations across various benchmarks demonstrate the effectiveness of the proposed method.

This paper questions the effectiveness of a modern predictive uncertainty quantification approach, called *evidential deep learning* (EDL), in which a single neural network model is trained to learn a meta distribution over the predictive distribution by minimizing a specific objective function. Despite their perceived strong empirical performance on downstream tasks, a line of recent studies by Bengs et al. identify limitations of the existing methods to conclude their learned epistemic uncertainties are unreliable, e.g., in that they are non-vanishing even with infinite data. Building on and sharpening such analysis, we 1) provide a sharper understanding of the asymptotic behavior of a wide class of EDL methods by unifying various objective functions; 2) reveal that the EDL methods can be better interpreted as an out-of-distribution detection algorithm based on energy-based-models; and 3) conduct extensive ablation studies to better assess their empirical effectiveness with real-world datasets. Through all these analyses, we conclude that even when EDL methods are empirically effective on downstream tasks, this occurs despite their poor uncertainty quantification capabilities. Our investigation suggests that incorporating model uncertainty can help EDL methods faithfully quantify uncertainties and further improve performance on representative downstream tasks, albeit at the cost of additional computational complexity.

It is known that neural networks have the problem of being over-confident when directly using the output label distribution to generate uncertainty measures. Existing methods mainly resolve this issue by retraining the entire model to impose the uncertainty quantification capability so that the learned model can achieve desired performance in accuracy and uncertainty prediction simultaneously. However, training the model from scratch is computationally expensive, and a trade-off might exist between prediction accuracy and uncertainty quantification. To this end, we consider a more practical post-hoc uncertainty learning setting, where a well-trained base model is given, and we focus on the uncertainty quantification task at the second stage of training. We propose a novel Bayesian uncertainty learning approach using the Dirichlet meta-model, which is effective and computationally efficient. Our proposed method requires no additional training data and is flexible enough to quantify different uncertainties and easily adapt to different application settings, including out-of-domain data detection, misclassification detection, and trustworthy transfer learning. Finally, we demonstrate our proposed meta-model approach's flexibility and superior empirical performance on these applications over multiple representative image classification benchmarks.

Gaussian processes (GPs) are used to make medical and scientific decisions, including in cardiac care and monitoring of carbon dioxide emissions. Notably, the choice of GP kernel is often somewhat arbitrary. In particular, uncountably many kernels typically align with qualitative prior knowledge (e.g. function smoothness or stationarity). But in practice, data analysts choose among a handful of convenient standard kernels (e.g. squared exponential). In the present work, we ask: Would decisions made with a GP differ under other, qualitatively interchangeable kernels? We show how to formulate this sensitivity analysis as a constrained optimization problem over a finite-dimensional space. We can then use standard optimizers to identify substantive changes in relevant decisions made with a GP. We demonstrate in both synthetic and real-world examples that decisions made with a GP can exhibit substantial sensitivity to kernel choice, even when prior draws are qualitatively interchangeable to a user.

In consequential decision-making applications, mitigating unwanted biases in machine learning models that yield systematic disadvantage to members of groups delineated by sensitive attributes such as race and gender is one key intervention to strive for equity. Focusing on demographic parity and equality of opportunity, in this paper we propose an algorithm that improves the fairness of a pre-trained classifier by simply dropping carefully selected training data points. We select instances based on their influence on the fairness metric of interest, computed using an infinitesimal jackknife-based approach. The dropping of training points is done in principle, but in practice does not require the model to be refit. Crucially, we find that such an intervention does not substantially reduce the predictive performance of the model but drastically improves the fairness metric. Through careful experiments, we evaluate the effectiveness of the proposed approach on diverse tasks and find that it consistently improves upon existing alternatives.

Bayesian decision theory provides an elegant framework for acting optimally under uncertainty when tractable posterior distributions are available. Modern Bayesian models, however, typically involve intractable posteriors that are approximated with, potentially crude, surrogates. This difficulty has engendered loss-calibrated techniques that aim to learn posterior approximations that favor high-utility decisions. In this paper, focusing on Bayesian neural networks, we develop methods for correcting approximate posterior predictive distributions encouraging them to prefer high-utility decisions. In contrast to previous work, our approach is agnostic to the choice of the approximate inference algorithm, allows for efficient test time decision making through amortization, and empirically produces higher quality decisions. We demonstrate the effectiveness of our approach through controlled experiments spanning a diversity of tasks and datasets.

Clinical researchers use disease progression models to understand patient status and characterize progression patterns from longitudinal health records. One approach for disease progression modeling is to describe patient status using a small number of states that represent distinctive distributions over a set of observed measures. Hidden Markov models (HMMs) and its variants are a class of models that both discover these states and make inferences of health states for patients. Despite the advantages of using the algorithms for discovering interesting patterns, it still remains challenging for medical experts to interpret model outputs, understand complex modeling parameters, and clinically make sense of the patterns. To tackle these problems, we conducted a design study with clinical scientists, statisticians, and visualization experts, with the goal to investigate disease progression pathways of chronic diseases, namely type 1 diabetes (T1D), Huntington's disease, Parkinson's disease, and chronic obstructive pulmonary disease (COPD). As a result, we introduce DPVis which seamlessly integrates model parameters and outcomes of HMMs into interpretable and interactive visualizations. In this article, we demonstrate that DPVis is successful in evaluating disease progression models, visually summarizing disease states, interactively exploring disease progression patterns, and building, analyzing, and comparing clinically relevant patient subgroups.

We propose a method to fuse posterior distributions learned from heterogeneous datasets. Our algorithm relies on a mean field assumption for both the fused model and the individual dataset posteriors and proceeds using a simple assign-and-average approach. The components of the dataset posteriors are assigned to the proposed global model components by solving a regularized variant of the assignment problem. The global components are then updated based on these assignments by their mean under a KL divergence. For exponential family variational distributions, our formulation leads to an efficient non-parametric algorithm for computing the fused model. Our algorithm is easy to describe and implement, efficient, and competitive with state-of-the-art on motion capture analysis, topic modeling, and federated learning of Bayesian neural networks.

Many modern data analyses benefit from explicitly modeling dependence structure in data -- such as measurements across time or space, ordered words in a sentence, or genes in a genome. A gold standard evaluation technique is structured cross-validation (CV), which leaves out some data subset (such as data within a time interval or data in a geographic region) in each fold. But CV here can be prohibitively slow due to the need to re-run already-expensive learning algorithms many times. Previous work has shown approximate cross-validation (ACV) methods provide a fast and provably accurate alternative in the setting of empirical risk minimization. But this existing ACV work is restricted to simpler models by the assumptions that (i) data across CV folds are independent and (ii) an exact initial model fit is available. In structured data analyses, both these assumptions are often untrue. In the present work, we address (i) by extending ACV to CV schemes with dependence structure between the folds. To address (ii), we verify -- both theoretically and empirically -- that ACV quality deteriorates smoothly with noise in the initial fit. We demonstrate the accuracy and computational benefits of our proposed methods on a diverse set of real-world applications.

Diabetes mellitus, commonly known as diabetes, is a chronic disease that often results in multiple complications. Risk prediction of diabetes complications is critical for healthcare professionals to design personalized treatment plans for patients in diabetes care for improved outcomes. In this paper, focusing on Type 2 diabetes mellitus (T2DM), we study the risk of developing complications after the initial T2DM diagnosis from longitudinal patient records. We propose a novel multi-task learning approach to simultaneously model multiple complications where each task corresponds to the risk modeling of one complication. Specifically, the proposed method strategically captures the relationships (1) between the risks of multiple T2DM complications, (2) between different risk factors, and (3) between the risk factor selection patterns, which assumes similar complications have similar contributing risk factors. The method uses coefficient shrinkage to identify an informative subset of risk factors from high-dimensional data, and uses a hierarchical Bayesian framework to allow domain knowledge to be incorporated as priors. The proposed method is favorable for healthcare applications because in addition to improved prediction performance, relationships among the different risks and among risk factors are also identified. Extensive experimental results on a large electronic medical claims database show that the proposed method outperforms state-of-the-art models by a significant margin. Furthermore, we show that the risk associations learned and the risk factors identified lead to meaningful clinical insights.

In unsupervised learning, dimensionality reduction is an important tool for data exploration and visualization. Because these aims are typically open-ended, it can be useful to frame the problem as looking for patterns that are enriched in one dataset relative to another. These pairs of datasets occur commonly, for instance a population of interest vs. control or signal vs. signal free recordings. However, there are few methods that work on sets of data as opposed to data points or sequences. Here, we present a probabilistic model for dimensionality reduction to discover signal that is enriched in the target dataset relative to the background dataset. The data in these sets do not need to be paired or grouped beyond set membership. By using a probabilistic model where some structure is shared amongst the two datasets and some is unique to the target dataset, we are able to recover interesting structure in the latent space of the target dataset. The method also has the advantages of a probabilistic model, namely that it allows for the incorporation of prior information, handles missing data, and can be generalized to different distributional assumptions. We describe several possible variations of the model and demonstrate the application of the technique to de-noising, feature selection, and subgroup discovery settings.

In federated learning problems, data is scattered across different servers and exchanging or pooling it is often impractical or prohibited. We develop a Bayesian nonparametric framework for federated learning with neural networks. Each data server is assumed to provide local neural network weights, which are modeled through our framework. We then develop an inference approach that allows us to synthesize a more expressive global network without additional supervision, data pooling and with as few as a single communication round. We then demonstrate the efficacy of our approach on federated learning problems simulated from two popular image classification datasets.

We consider the problem of aggregating models learned from sequestered, possibly heterogeneous datasets. Exploiting tools from Bayesian nonparametrics, we develop a general meta-modeling framework that learns shared global latent structures by identifying correspondences among local model parameterizations. Our proposed framework is model-independent and is applicable to a wide range of model types. After verifying our approach on simulated data, we demonstrate its utility in aggregating Gaussian topic models, hierarchical Dirichlet process based hidden Markov models, and sparse Gaussian processes with applications spanning text summarization, motion capture analysis, and temperature forecasting.

The promise of augmenting accurate predictions provided by modern neural networks with well-calibrated predictive uncertainties has reinvigorated interest in Bayesian neural networks. However, model selection---even choosing the number of nodes---remains an open question. Poor choices can severely affect the quality of the produced uncertainties. In this paper, we explore continuous shrinkage priors, the horseshoe, and the regularized horseshoe distributions, for model selection in Bayesian neural networks. When placed over node pre-activations and coupled with appropriate variational approximations, we find that the strong shrinkage provided by the horseshoe is effective at turning off nodes that do not help explain the data. We demonstrate that our approach finds compact network structures even when the number of nodes required is grossly over-estimated. Moreover, the model selection over the number of nodes does not come at the expense of predictive or computational performance; in fact, we learn smaller networks with comparable predictive performance to current approaches. These effects are particularly apparent in sample-limited settings, such as small data sets and reinforcement learning.

Type 2 diabetes mellitus (T2DM) is a chronic disease that usually results in multiple complications. Early identification of individuals at risk for complications after being diagnosed with T2DM is of significant clinical value. In this paper, we present a new data-driven predictive approach to predict when a patient will develop complications after the initial T2DM diagnosis. We propose a novel survival analysis method to model the time-to-event of T2DM complications designed to simultaneously achieve two important metrics: 1) accurate prediction of event times, and 2) good ranking of the relative risks of two patients. Moreover, to better capture the correlations of time-to-events of the multiple complications, we further develop a multi-task version of the survival model. To assess the performance of these approaches, we perform extensive experiments on patient level data extracted from a large electronic health record claims database. The results show that our new proposed survival analysis approach consistently outperforms traditional survival models and demonstrate the effectiveness of the multi-task framework over modeling each complication independently.

Objective automated affect analysis systems can be applied to quantify the progression of symptoms in neurodegenerative diseases such as Parkinson's Disease (PD). PD hampers the ability of patients to emote by decreasing the mobility of their facial musculature, a phenomenon known as ``facial masking.'' In this work, we focus on building a system that can predict an accurate score of active facial expressivity in people suffering from Parkinson's disease using features extracted from both video and audio. An ideal automated system should be able to mimic the ability of human experts to take into account contextual information while making these predictions. For example, patients exhibit different emotions with varying intensities when speaking about positive and negative experiences. We utilize a hierarchical Bayesian neural network framework to enable the learning of model parameters that subtly adapt to pre-defined notions of context, such as the gender of the patient or the valence of the expressed sentiment. We evaluate our formulation on a dataset of 772 20-second video clips of Parkinson's disease patients and demonstrate that training a context-specific hierarchical Bayesian framework yields an improvement in model performance in both multiclass classification and regression settings compared to baseline models trained on all data pooled together.

Bayesian Neural Networks (BNNs) have recently received increasing attention for their ability to provide well-calibrated posterior uncertainties. However, model selection—even choosing the number of nodes—remains an open question. Recent work has proposed the use of a horseshoe prior over node pre-activations of a Bayesian neural network, which effectively turns off nodes that do not help explain the data. In this work, we propose several modeling and inference advances that consistently improve the compactness of the model learned while maintaining predictive performance, especially in smaller-sample settings including reinforcement learning.

Building robust classifiers trained on data susceptible to group or subject-specific variations is a challenging pattern recognition problem. We develop hierarchical Bayesian neural networks to capture subject-specific variations and share statistical strength across subjects. Leveraging recent work on learning Bayesian neural networks, we build fast, scalable algorithms for inferring the posterior distribution over all network weights in the hierarchy. We also develop methods for adapting our model to new subjects when a small number of subject-specific personalization data is available. Finally, we investigate active learning algorithms for interactively labeling personalization data in resource-constrained scenarios. Focusing on the problem of gesture recognition where inter-subject variations are commonplace, we demonstrate the effectiveness of our proposed techniques. We test our framework on three widely used gesture recognition datasets, achieving personalization performance competitive with the state-of-the-art.

Buoyed by the success of deep multilayer neural networks, there is renewed interest in scalable learning of Bayesian neural networks. Here, we study algorithms that utilize recent advances in Bayesian inference to efficiently learn distributions over network weights. In particular, we focus on recently proposed assumed density filtering based methods for learning Bayesian neural networks -- Expectation and Probabilistic backpropagation. Apart from scaling to large datasets, these techniques seamlessly deal with non-differentiable activation functions and provide parameter (learning rate, momentum) free learning. In this paper, we first rigorously compare the two algorithms and in the process develop several extensions, including a version of EBP for continuous regression problems and a PBP variant for binary classification. Next, we extend both algorithms to deal with multiclass classification and count regression problems. On a variety of diverse real world benchmarks, we find our extensions to be effective, achieving results competitive with the state-of-the-art.

We explore recently proposed Bayesian nonparametric models of image partitions, based on spatially dependent Pitman-Yor processes. These models are attractive because they adapt to images of varying complexity, successfully modeling uncertainty in the structure and scale of human segmentations of natural scenes. By developing substantially improved inference and learning algorithms, we achieve performance comparable to state-of-the-art methods. For learning, we show how the Gaussian process (GP) covariance functions underlying these models can be calibrated to accurately match the statistics of example human segmentations. For inference, we develop a stochastic search-based algorithm which is substantially less susceptible to local optima than conventional variational methods. Our approach utilizes the expectation propagation algorithm to approximately marginalize latent GPs, and a low rank covariance representation to improve computational efficiency. Experiments with two benchmark datasets show that our learning and inference innovations substantially improve segmentation accuracy. By hypothesizing multiple partitions for each image, we also take steps towards capturing the variability of human scene interpretations.

We develop a method for discovering the parts of an articulated object from aligned meshes capturing various three-dimensional (3D) poses. We adapt the distance dependent Chinese restaurant process (ddCRP) to allow nonparametric discovery of a potentially unbounded number of parts, while simultaneously guaranteeing a spatially connected segmentation. To allow analysis of datasets in which object instances have varying shapes, we model part variability across poses via affine transformations. By placing a matrix normal-inverse-Wishart prior on these affine transformations, we develop a ddCRP Gibbs sampler which tractably marginalizes over transformation uncertainty. Analyzing a dataset of humans captured in dozens of poses, we infer parts which provide quantitatively better motion predictions than conventional clustering methods.

The distance dependent Chinese restaurant process (ddCRP) was recently introduced to accommodate random partitions of non-exchangeable data. The ddCRP clusters data in a biased way: each data point is more likely to be clustered with other data that are near it in an external sense. This paper examines the ddCRP in a spatial setting with the goal of natural image segmentation. We explore the biases of the spatial ddCRP model and propose a novel hierarchical extension better suited for producing "human-like" segmentations. We then study the sensitivity of the models to various distance and appearance hyperparameters, and provide the first rigorous comparison of nonparametric Bayesian models in the image segmentation domain. On unsupervised image segmentation, we demonstrate that similar performance to existing nonparametric Bayesian models is possible with substantially simpler models and algorithms.

The ability to identify directional interactions that occur among multiple neurons in the brain is crucial to an understanding of how groups of neurons cooperate in order to generate specific brain functions. However, an optimal method of assessing these interactions has not been established. Granger causality has proven to be an effective method for the analysis of the directional interactions between multiple sets of continuous-valued data, but cannot be applied to neural spike train recordings due to their discrete nature. This paper proposes a point process framework that enables Granger causality to be applied to point process data such as neural spike trains. The proposed framework uses the point process likelihood function to relate a neuron's spiking probability to possible covariates, such as its own spiking history and the concurrent activity of simultaneously recorded neurons. Granger causality is assessed based on the relative reduction of the point process likelihood of one neuron obtained excluding one of its covariates compared to the likelihood obtained using all of its covariates. The method was tested on simulated data, and then applied to neural activity recorded from the primary motor cortex (MI) of a Felis catus subject. The interactions present in the simulated data were predicted with a high degree of accuracy, and when applied to the real neural data, the proposed method identified causal relationships between many of the recorded neurons. This paper proposes a novel method that successfully applies Granger causality to point process data, and has the potential to provide unique physiological insights when applied to neural spike trains.

Navigating through unknown outdoor environments requires a robot to be able to see and model the far field terrain. In recent years this problem of seeing beyond reliable stereo readings into the far field has gained attention. Many proposed solutions involve using near field obstacle and ground plane regions labeled using stereo, to learn models which classify far field image regions. In this work we offer an alternative which exploits coherent image regions as determined by image segmentation to propagate obstacle and ground labels from the near and mid field to the image far field. Rather than relying on local features to classify individual pixels we model and compare appearance across the whole segment. New labels are determined by proximity in both image space and appearance space. Since both traversable and non-traversable surfaces can vary in appearance across the image, our approach has the advantage that each labeled segment acts as a distinct appearance model, which allows us to label similar neighbours. We evaluate our system using a publicly available dataset and compare its performance to a typical learning-based near-to-far labeling scheme.

In this paper, we present a methodology for automatic geomorphic mapping of planetary surfaces that incorporates machine-learning techniques. Our application transforms remotely sensed topographic data gathered by orbiting satellites into semantically meaningful maps of landforms; such maps are valuable research tools for planetary science. As topographic data become increasingly available, the ability to derive geomorphic maps efficiently is becoming essential. In our proposed framework, the mapping is achieved by means of scene segmentation followed by supervised classification of segments. The two mapping steps use different sets of features derived from digital elevation models of planetary surfaces; selection of appropriate features is discussed. Using a particular set of terrain attributes relevant to annotating cratered terrain on Mars, we investigate the design choices for both segmentation and classification components. The segmentation assessment includesK-means-based agglomerative segmentation and watershed-based segmentation. The classification assessment includes three supervised learning algorithms: Naive Bayes, Bagging with decision trees, and support-vector machines (SVMs); segments are classified into the following landforms: crater floors, crater walls (concave and convex), ridges (concave and convex), and intercrater plains. The method is applied to six test sites on Mars. The analysis of the results shows that a combination ofK-means-based agglomerative segmentation and either SVM with a quadratic kernel or Bagging with C4.5 yields best maps. The presented framework can be adopted to generate geomorphic maps of sites on Earth.

The expectation maximization algorithm is a popular approach to learning Gaussian mixture models from unlabeled data. In addition to the unlabeled data, in many applications, additional sources of information such as a-priori knowledge of mixing proportions are also available. We present a weakly supervised approach, in the form of a penalized expectation maximization algorithm that uses a-priori knowledge to guide the model training process. The algorithm penalizes those models whose predicted mixing proportions have high divergence from the a-priori mixing proportions. We also present an extension to incorporate both labeled and unlabeled data in a semi-supervised setting. Systematic evaluations on several publicly available datasets show that the proposed algorithms outperforms the expectation maximization algorithm. The performance gains are particularly significant when the amount of unlabeled data is limited and in the presence of noise.

Segmentation, or partitioning images into internally homogeneous regions, is an important first step in many Computer Vision tasks. In this paper, we attack the segmentation problem using an ensemble of low cost image segmentations. These segmentations are reconciled by applying recent techniques from the consensus clustering literature which exploit a Non-negative Matrix Factorization (NMF) framework. We describe extensions to these methods that scale them for large images and also incorporate smoothness constraints. This framework allows us to uniformly and easily combine segmentations from different algorithms or feature modalities. We then demonstrate that popular bottom up image segmentation algorithms, Mean Shift and Efficient Graph Based segmentation, perform no better than our simple combination of multiple image segmentations derived from k-means clustering (of various feature spaces) or from “naive” RGB quantizations. The algorithms are evaluated on the Berkeley image segmentation dataset.