Enhancing trustworthiness in probabilistic programming: systematic approaches for robust and accurate inference

Probabilistic programming simplifies the encoding of statistical models as straightforward programs. At its core, it employs an inference algorithm which automate the model inference, allowing developers to focus on model creation. Its simplicity has led to its growing application in critical areas such as autonomous driving, privacy modeling, computer networks, and pandemic prediction. However, the flexibility of probabilistic modeling and the scalability to large datasets come at a price of trustworthiness: the approximate inference algorithms used by many existing probabilistic programming systems may produce inaccurate results; also the collected data often contain noise, which can violate the model’s assumption and cause large deviations in the results. Trustworthiness, therefore, has two key properties: accuracy, to ensure results close to the true underlying distribution, and robustness, to ensure reliable results amidst data noise. This dissertation introduces a systematic approach, composed of multiple probabilistic programming systems throughout the probabilistic programming computation stack, to analyze and enhance the trustworthiness of probabilistic programs. We present the results of two-pronged investigation: first, we identify several trustworthiness challenges of the current practice of probabilistic programming algorithms. The examination is supported by the presentation of ASTRA, an experimental testbed for evaluating the robustness of probabilistic programs against data noise, and SixthSense, a system aiding developers in debugging convergence issues in sampling-based approximate inference algorithms. The second part of the dissertation introduces AQUA, a novel quantized inference algorithm which can achieve better accuracy than existing approximate inference algorithms and scales better than exact inference. Then, the dissertation advances the domain of probabilistic reasoning by moving beyond the conventional focus on computing a single posterior distribution. It presents AURA, an abstract interpretation which provides precise, soundly guaranteed bounds on posterior distributions when they are subjected an infinite set of data perturbations.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2024

Author dc:creator

• Huang, Zixin

Contributors dc:contributor

• Misailovic, Sasa
• Marinov, Darko
• Mitra, Sayan
• Kang, Eunsuk

• Probabilistic Programming
• Quantized Inference
• Robustness
• Abstract Interpretation
• Program Analysis
• Machine Learning

Statement dc:rights

• Copyright 2024 Zixin Huang

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/124259

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F124259