Publications

Probabilistic Models (PMs) provide a powerful framework for representing large multivariate probability distributions, but as their expressiveness increases, inference tasks become computationally intractable. In the era of big data, the demand for complex models capable of effectively representing and reasoning over large-scale data has increased significantly. While models like probabilistic circuits allow tractable inference for specific query types, handling a broader range of queries remains challenging. Exact inference in PMs is generally NP-hard, and approximate methods, though viable, often involve trade-offs between accuracy and efficiency, leading to unreliable estimates. Achieving near-optimal solutions typically demands substantial computational resources, posing a barrier to scalability.This dissertation presents novel algorithms for key inference tasks in PMs, including Most Probable Explanation (MPE), Constrained Most Probable Explanation (CMPE), and Marginal Maximum A Posteriori (MMAP) inference. Our methods generalize across different Probabilistic Models (PMs), such as Probabilistic Graphical Models (PGMs), Probabilistic Circuits (PCs), and Neural Autoregressive Models (NAMs), enabling efficient and scalable probabilistic reasoning. Additionally, we propose a neurosymbolic framework that integrates Dependency Networks (DNs) with deep learning architectures to enhance multi-label classification. Specifically, this dissertation makes the following contributions:We develop a self-supervised learning framework for neural network-based solvers to answer predefined MPE and MMAP queries over PCs. Our approach introduces a scalable loss function with a computation cost that scales linearly with the size of the PC, enabling the training of neural solvers that match or surpass state-of-the-art solvers in solution quality while reducing inference time from seconds to microseconds.We extend inference capabilities beyond predefined queries, allowing neural solvers to answer arbitrary MPE queries over PCs, PGMs, and NAMs. To achieve this, we introduce a novel dual-network learning framework and an enhanced inference scheme that updates neural network parameters at test time, leading to improved solution quality.We eliminate the need for costly inference-time optimization in neural solvers for arbitrary MPE queries over PGMs by enhancing query encoding, which improves solution quality through richer embeddings. Additionally, we introduce two novel methods for discretizing continuous neural network outputs, further enhancing solution quality.We extend neural solvers to constrained inference, specifically the Constrained Most Probable Explanation (CMPE) task. We develop a self-supervised CMPE solver with a loss function satisfying the consistent loss property, ensuring alignment with the optimal CMPE solution—unlike existing loss functions for constrained optimization, which lack this guarantee.We propose advanced inference schemes for MPE in neurosymbolic models for multi-label classification, specifically improving inference in Deep Dependency Networks (DDNs). While DDNs offer efficient training and an intuitive loss function for multi-label classification, they traditionally rely on Gibbs sampling, which limits inference accuracy and efficiency. To address this limitation, we introduce novel inference techniques based on local search and integer linear programming (ILP), facilitating more accurate and efficient computation of the most probable label assignments.

Diffusion processes on networks are central to social network analysis and data mining, where influence maximization provides a canonical framework for identifying influential nodes. However, many real-world diffusion processes are competitive and reversible, making classical irreversible influence maximization models inadequate. The positional Moran process captures such dynamics by modeling stochastic competition with position-dependent fitness and motivates the fixation maximization problem: selecting an initial active set that maximizes the probability of eventual population-wide dominance. Existing methods rely on static structural heuristics or simulation-intensive greedy search. These approaches either ignore long-horizon evolutionary effects or scale poorly because they require repeated stochastic simulations. This limitation raises a fundamental question: Can we learn scalable active node selection policies that capture the evolutionary dynamics governing fixation? We answer this question with LEAP, a deep reinforcement learning framework for amortized fixation maximization under the positional Moran process. LEAP formulates active node selection as a Markov decision problem and learns fixation-aware representations that combine graph structure with mutant invasion and stability signals. Its fixation-modulated message passing mechanism models selection pressure through learned fitness contrasts, while fixation-weighted pooling constructs a global evolutionary representation. Experiments on real-world networks show that LEAP matches the solution quality of simulation-based greedy methods with approximately five orders of magnitude faster inference and outperforms all baselines in at least two-thirds of experimental settings. These results show that deep reinforcement learning can optimize long-horizon competitive diffusion when equipped with fixation-aware mechanisms that capture evolutionary dynamics.

Probabilistic graphical models (PGMs) compactly encode structured joint distributions, but Most Probable Explanation (MPE), which finds the highest-probability assignment to unobserved variables, remains computationally intractable in large, densely structured models. Anytime systematic search, such as branch-and-bound (B&B), is a popular approach, but its efficiency depends on tight admissible upper bounds that are often prohibitively expensive to compute using classical methods like Mini-Bucket Elimination (MBE). We introduce Learning to Bound (L2B), a data-driven framework that amortizes bound computation cost by training a neural predictor. L2B targets two key properties required for effective search bounds: admissibility, which is achieved via statistical calibration techniques, and progressive tightening as evidence accumulates, which is enforced through structural architectural constraints that ensure monotonicity by construction. Across challenging benchmarks, L2B delivers tighter, monotonic bounds at orders-of-magnitude lower cost on every PGM family, and yields measurable improvements when integrated with branch-and-bound on high-treewidth instances.

Human-in-the-loop methods leverage human feedback to enhance machine learning and AI. Manual review of outputs can correct errors, identify model weaknesses, or expand labels to broaden model capabilities. Feedback collection methods range from simple flagging of outputs as correct or incorrect to more complex feature-level adjustments or natural language interpretations. This article presents a user study evaluating changes in user performance over time and explores the tradeoff between feedback quality and human effort. We compare four interactive input methods for reviewing and correcting outcomes in object detection and activity recognition in videos. Our findings indicate that while some complex input methods, such as free-text, require more time, the quality and impact of their feedback on model accuracy often surpass those of simpler methods that require less effort. However, more effort does not always lead to better-quality feedback, especially when aiming to improve the model. Our VLM experiments show that the most accurate models were trained using detailed natural language feedback or precise word-level corrections, while simple yes/no judgments also led to solid performance at a much lower annotation cost.

Single-cell RNA sequencing provides a powerful view of cellular heterogeneity, but its sparsity and dropout noise remain major obstacles for recovering biologically meaningful gene expression programs and for downstream analyses that depend on reliable expression measurements. Ligand–receptor-based cell–cell communication inference is such analysis, missing ligand or receptor expression can cause substantial false negatives in sparse single-cell data. Here, we present CoLa-VAE, a cell–cell communication-aware variational autoencoder that jointly learns latent representations and denoised expression profiles by incorporating ligand–receptor-derived communication topology through dynamic graph Laplacian regularization. Rather than treating denoising as a secondary output of representation learning, CoLa-VAE uses denoised expression to iteratively refine communication estimates and uses the resulting communication structure to guide both latent organization and expression reconstruction. In addition to improving latent space organization and producing robust denoised expression matrices, CoLa-VAE-denoised matrices also improved downstream biological analyses, including the detection of robust differential cell–cell communication programs, mitigation of batch-associated variation and enhanced spatial transcriptomic deconvolution when spatially constrained communication structure was incorporated. Together, these results establish CoLa-VAE as a communication-guided denoising and representation learning framework that recovers biologically meaningful expression signals from sparse single-cell and spatial transcriptomic data, enabling more sensitive and reliable downstream analysis.Competing Interest StatementThe authors have declared no competing interest.

Most Probable Explanation (MPE) inference in Probabilistic Graphical Models (PGMs) is a fundamental yet computationally challenging problem arising in domains such as diagnosis, planning, and structured prediction. In many practical settings, the graphical model remains fixed while inference must be performed repeatedly for varying evidence patterns. Stochastic Local Search (SLS) algorithms scale to large models but rely on myopic best-improvement rule that prioritizes immediate likelihood gains and often stagnate in poor local optima. Heuristics such as Guided Local Search (GLS+) partially alleviate this limitation by modifying the search landscape, but their guidance cannot be reused effectively across multiple inference queries on the same model. We propose a neural amortization framework for improving local search in this repeated-query regime. Exploiting the fixed graph structure, we train an attention-based network to score local moves by predicting their ability to reduce Hamming distance to a near-optimal solution. Our approach integrates seamlessly with existing local search procedures, using this signal to balance short-term likelihood gains with long-term promise during neighbor selection. We provide theoretical intuition linking distance-reducing move selection to improved convergence behavior, and empirically demonstrate consistent improvements over SLS and GLS+ on challenging high-treewidth benchmarks in the amortized inference setting.

We introduce learning to condition (L2C), a scalable, data-driven framework for accelerating Most Probable Explanation (MPE) inference in Probabilistic Graphical Models (PGMs), a fundamentally intractable problem. L2C trains a neural network to score variable-value assignments based on their utility for conditioning, given observed evidence. To facilitate supervised learning, we develop a scalable data generation pipeline that extracts training signals from the search traces of existing MPE solvers. The trained network serves as a heuristic that integrates with search algorithms, acting as a conditioning strategy prior to exact inference or as a branching and node selection policy within branch-and-bound solvers. We evaluate L2C on challenging MPE queries involving high-treewidth PGMs. Experiments show that our learned heuristic significantly reduces the search space while maintaining or improving solution quality over state-of-the-art methods.

Our paper builds on the recent trend of using neural networks trained with self-supervised or supervised learning to solve the Most Probable Explanation (MPE) task in discrete graphical models. At inference time, these networks take an evidence assignment as input and generate the most likely assignment for the remaining variables via a single forward pass. We address two key limitations of existing approaches: (1) the inability to fully exploit the graphical model’s structure and parameters, and (2) the suboptimal discretization of continuous neural network outputs. Our approach embeds model structure and parameters into a more expressive feature representation, significantly improving  performance. Existing methods rely on standard thresholding, which often yields suboptimal results due to the non-convexity of the loss function. We introduce two methods to overcome discretization challenges: (1) an external oracle-based approach that infers uncertain variables using additional evidence from confidently predicted ones, and (2) a technique that identifies and selects the highest-scoring discrete solutions near the continuous output. Experimental results on various probabilistic models demonstrate the effectiveness and scalability of our approach, highlighting its practical impact.

Following step-by-step procedures is an essential component of various activities carried out by individuals in their daily lives. These procedures serve as a guiding framework that helps to achieve goals efficiently, whether it is assembling furniture or preparing a recipe. However, the complexity and duration of procedural activities inherently increase the likelihood of making errors. Understanding such procedural activities from a sequence of frames is a challenging task that demands an accurate interpretation of visual information and the ability to reason about the structure of the activity. To this end, we collect a new egocentric 4D dataset CaptainCook4D comprising 384 recordings (94.5 hours) of people performing recipes in real kitchen environments. This dataset consists of two distinct types of activities: one in which participants adhere to the provided recipe instructions and another in which they deviate and induce errors. We provide 5.3K step annotations and 10K finegrained action annotations and benchmark the dataset for the following tasks: error recognition, multi-step localization and procedure learning. https://captaincook4d.github.io/captain-cook/

We propose a self-supervised learning approach for solving the following constrained optimization task in log-linear models or Markov networks. Let f and g be two log-linear models defined over the sets X and Y of random variables respectively. Given an assignment x to all variables in X (evidence) and a real number q, the constrained most-probable explanation (CMPE) task seeks to find an assignment y to all variables in Y such that f(x,y) is maximized and g(x,y)≤q. In our proposed self-supervised approach, given assignments x to X (data), we train a deep neural network that learns to output near-optimal solutions to the CMPE problem without requiring access to any pre-computed solutions. The key idea in our approach is to use first principles and approximate inference methods for CMPE to derive novel loss functions that seek to push infeasible solutions towards feasible ones and feasible solutions towards optimal ones. We analyze the properties of our proposed method and experimentally demonstrate its efficacy on several benchmark problems.

We present a unified framework called deep dependency networks (DDNs) that combines dependency networks and deep learning architectures for multi-label classification, with a particular emphasis on image and video data. The primary advantage of dependency networks is their ease of training, in contrast to other probabilistic graphical models like Markov networks. In particular, when combined with deep learning architectures, they provide an intuitive, easy-to-use loss function for multi-label classification. A drawback of DDNs compared to Markov networks is their lack of advanced inference schemes, necessitating the use of Gibbs sampling. To address this challenge, we propose novel inference schemes based on local search and integer linear programming for computing the most likely assignment to the labels given observations. We evaluate our novel methods on three video datasets (Charades, TACoS, Wetlab) and three image datasets (MS-COCO, PASCAL VOC, NUS-WIDE), comparing their performance with (a) basic neural architectures and (b) neural architectures combined with Markov networks equipped with advanced inference and learning techniques. Our results demonstrate the superiority of our new DDN methods over the two competing approaches.

We put forward an augmented reality (AR) system using artificial intelligence (AI) to guide users through real-world procedures. The system uses the cameras of the headset to recognize objects and their positions in real time. It constructs a model to predict: (i) the task being completed, (ii) the step of the task they are on, and (iii) whether they have made any errors. By updating this model in real-time as the user completes the task, the system automatically updates the instructions and cues provided to the user. This system represents a step towards ubiquitous task guidance via AR.

We consider the following video activity recognition (VAR) task: given a video, infer the set of activities being performed in the video and assign each frame to an activity. Although VAR can be solved accurately using existing deep learning techniques, deep networks are neither interpretable nor explainable and as a result their use is problematic in high stakes decision-making applications (e.g., in healthcare, experimental Biology, aviation, law, etc.). In such applications, failure may lead to disastrous consequences and therefore it is necessary that the user is able to either understand the inner workings of the model or probe it to understand its reasoning patterns for a given decision. We address these limitations of deep networks by proposing a new approach that feeds the output of a deep model into a tractable, interpretable probabilistic model called a dynamic conditional cutset network that is defined over the explanatory and output variables and then performing joint inference over the combined model. The two key benefits of using cutset networks are: (a) they explicitly model the relationship between the output and explanatory variables and as a result the combined model is likely to be more accurate than the vanilla deep model and (b) they can answer reasoning queries in polynomial time and as a result they can derive meaningful explanations by efficiently answering explanation queries. We demonstrate the efficacy of our approach on two datasets, Textually Annotated Cooking Scenes (TACoS), and wet lab, using conventional evaluation measures such as the Jaccard Index and Hamming Loss, as well as a human-subjects study.

Following step-by-step procedures is an essential component of various activities carried out by individuals in their everyday lives. These procedures serve as a guiding framework that helps achieve goals efficiently, whether assembling furniture or preparing a recipe. However, the complexity and duration of procedural activities inherently increase the likelihood of making errors. Understanding such procedural activities from a sequence of frames is a challenging task that demands an accurate interpretation of visual information and an ability to reason about the structure of the activity. To this end, we collected a new ego-centric 4D dataset comprising 380 recordings (90 hrs) of people performing recipes in kitchen environments. This dataset consists of two distinct activity types: one in which participants adhere to the provided recipe instructions and another where they deviate and induce errors. We provide 5K step annotations and 10K fine-grained action annotations for 20% of the collected data and benchmark it on two tasks: error detection and procedure learning.

We propose a simple approach which combines the strengths of probabilistic graphical models and deep learning architectures for solving the multi-label classification task, focusing specifically on image and video data. First, we show that the performance of previous approaches that combine Markov Random Fields with neural networks can be modestly improved by leveraging more powerful methods such as iterative join graph propagation, integer linear programming, and  regularization-based structure learning. Then we propose a new modeling framework called deep dependency networks, which augments a dependency network, a model that is easy to train and learns more accurate dependencies but is limited to Gibbs sampling for inference, to the output layer of a neural network. We show that despite its simplicity, jointly learning this new architecture yields significant improvements in performance over the baseline neural network. In particular, our experimental evaluation on three video activity classification datasets: Charades, Textually Annotated Cooking Scenes (TACoS), and Wetlab, and three multi-label image classification datasets: MS-COCO, PASCAL VOC, and NUS-WIDE show that deep dependency networks are almost always superior to pure neural architectures that do not use dependency networks.

In this paper, a kernelized version of the random vector functional link network is proposed for multi-label classification. This classifier uses pseudo-inverse to find output weights of the network. As pseudo-inverse is non-iterative in nature, it requires less fine-tuning to train the network. Kernelization of RVFL makes it robust and stable as no need to tune the number of neuron in the enhancement layer. A threshold function is used with a kernelized random vector functional link network to make it suitable for multi-label learning problems. Experiments performed on three benchmark multi-label datasets bibtex, emotions, and scene shows that proposed classifier outperforms various the existing multi-label classifiers.