Proceedings of the Thirty-Third Conference on Uncertainty in Artificial Intelligence, UAI 2017, Sydney, Australia, August 11-15, 2017

UAI 2017 · DBLP

UAI 2017 Proceedings Cover

Overview

The 33rd Conference on Uncertainty in Artificial Intelligence (UAI 2017), held in Sydney, Australia, from August 11–15, served as a premier venue for the latest advancements in probabilistic reasoning, machine learning, and decision-making under uncertainty. This volume, containing 87 peer-reviewed papers selected from 282 submissions, offers a comprehensive snapshot of a field in transition. In 2017, the UAI community was actively working to integrate the representational power of deep learning with the rigorous, interpretable frameworks of Bayesian inference.

Core Challenges in Probabilistic AI

The central mission of UAI is the development of systems that can reason and act effectively in the face of incomplete, noisy, or contradictory information. During this period, several critical tensions defined the research agenda:

  1. Scalability and Big Data: While deep learning thrived in "Big Data" regimes, traditional Bayesian methods often struggled with computational complexity. Researchers sought ways to make probabilistic models as scalable as their neural counterparts.
  2. Causal Inference: Moving beyond statistical correlation to understand the underlying causal mechanisms is vital for science and policy. The proceedings reflect a push to infer these structures from observational data alone.
  3. Reliability, Fairness, and Robustness: As AI systems moved into sensitive domains like healthcare and criminal justice, ensuring that automated decisions were not only accurate but also fair and resilient to "noisy" or biased training data became a priority.
  4. The Labeling Bottleneck: The high cost of human-annotated data remained a significant barrier. This led to a surge in interest in "weak supervision" and programmatic data generation.

Methodological Innovations

The papers in this volume showcase a diverse toolkit of algorithmic breakthroughs designed to address the aforementioned challenges.

Scaling Inference

A major focus was advancing Variational Inference (VI) to handle non-conjugate models and deep structures. For instance, researchers introduced Deep Gaussian Processes for tasks like inverse reinforcement learning, providing a more flexible way to model complex reward functions. Additionally, improvements in Markov Chain Monte Carlo (MCMC) methods, such as "minibatch" acceptance tests for Metropolis-Hastings, allowed these theoretically grounded sampling techniques to be applied to massive datasets that were previously inaccessible.

Causal Discovery and Structure Learning

The proceedings highlight the use of advanced computational tools, including Boolean satisfiability (SAT) solvers and algebraic equivalence classes, to infer causal structures. These methods allow for discovery under weaker assumptions than traditional algorithms, making them more applicable to real-world scenarios where data may be sparse or incomplete.

Weak Supervision and Data Programming

A significant theme, underscored by keynote speaker Christopher Ré, was the shift toward weak supervision. Rather than relying solely on manual annotation, the "Snorkel" framework and similar approaches allow researchers to programmatically label data using heuristics or existing knowledge bases. This "data programming" paradigm helps bypass the labeling bottleneck that often stalls complex model development.

Benchmarks and Applications

While standard image recognition benchmarks like MNIST, CIFAR-10, and SVHN remained prevalent for testing new architectures, the 2017 proceedings featured a notable expansion into high-stakes real-world domains.

A portion of the UAI 2017 Program Committee, reflecting the global and institutional diversity of the conference.
  • Healthcare and Medicine: Researchers applied probabilistic models to electronic health records and ICU monitoring data, addressing critical clinical tasks such as determining the optimal time to wean patients off mechanical ventilation.
  • Biological Sciences: Keynote speaker Terry Speed highlighted the role of uncertainty in single-cell omics and nanopore sequencing, where noise and variability are inherent to the data generation process.
  • Computer Vision and Tracking: Beyond standard classification, the proceedings included work on image partitioning using the Berkeley Segmentation Dataset and specialized tasks like tracking lion behavior using probabilistic programming.

Key Insights and Award-Winning Contributions

The conference recognized several papers that pushed the boundaries of the field:

  • Fairness in Policy Selection: The Best Paper Award, "Importance Sampling for Fair Policy Selection," demonstrated that standard importance sampling techniques—often used to evaluate new policies from historical data—can introduce systematic biases. The authors argued that fairness must be an intrinsic part of the estimator's design.
  • Limits of Probabilistic Logic: The Best Student Paper explored the expressive limits of real-valued probabilistic logic programs. This theoretical work helps explain why certain types of "rule-based" AI can become unexpectedly complex and difficult to manage.
  • Holographic Model States: One intriguing proposal viewed features in deep networks as partial derivatives of the output. This "holographic" perspective suggests that a simple linear model can sometimes replicate the behavior of a much more complex network, offering a path toward better model interpretability.

Results and Future Outlook

With a competitive 31% acceptance rate, UAI 2017 maintained high standards for technical rigor. The results presented were significant: the PartiMax algorithm for object tracking demonstrated it could maintain 80% of its original performance while processing only 10% of the pixel boxes, a major win for efficiency. Furthermore, the AutoGP framework broke the 1% error-rate barrier for Gaussian Process models on the MNIST8M dataset, proving that Bayesian methods can compete at the highest levels of scale and accuracy.

Looking forward, the proceedings suggest that the "black box" nature of deep learning remains a primary hurdle. Future work will likely focus on enhancing the interpretability of hybrid models, tackling the complexity of causal discovery in messy real-world time series, and reducing the computational overhead of uncertainty quantification for edge and real-time applications.

Summary updated

BibTeX
@inproceedings{2017proceedings,
  title = {{Proceedings of the Thirty-Third Conference on Uncertainty in Artificial Intelligence, UAI 2017, Sydney, Australia, August 11-15, 2017}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2017},
}
Figure