Bayesian Networks — Artificial Intelligence for Research, Analytics, and Reasoning
Stefan Conrady, Managing Partner, Bayesia USA
“Currently, Bayesian Networks have become one of the most complete, self-sustained and coherent formalisms used for knowledge acquisition, representation and application through computer systems.” (Bouhamed et al., 2015)
In this workshop, we illustrate how scientists in many fields of study — rather than only computer scientists — can employ Bayesian networks as a very practical form of Artificial Intelligence for exploring complex problems. We present the remarkably simple theory behind Bayesian networks and then demonstrate how to utilize them for research and analytics tasks with the BayesiaLab software platform. More specifically, we illustrate supervised and unsupervised machine learning algorithms for knowledge discovery in high-dimensional domains.
Also, while Artificial Intelligence is commonly associated with another buzzword, “Big Data,” we show that Bayesian networks can bring Artificial Intelligence to problems for which we possess little or no data. Here, expert knowledge modeling is critical, and we describe how even a minimal amount of expertise can serve as a basis for robust reasoning under uncertainty with Bayesian networks.
Finally, theoretical expert knowledge remains absolutely mandatory for performing causal inference with non-experimental data. However, we show how the number causal assumptions can be substantially and conveniently reduced by using machine-learned Bayesian network as the model.
Seminar Program Part I (Approx. 60 minutes)
Motivations:
- The Promise, the Peril, and the Limitations of Artificial Intelligence
- Human Cognitive Limitations & Biases in Reasoning
Objectives:
- Human-Machine Teaming
- Practical Artificial Intelligence for Here & Now
Background: A Conceptual Map of Analytic Modeling and Reasoning
- X — Inference Type: Probabilistic vs. Deterministic
- Y — Model Purpose: Observational vs. Causal Inference
- Z — Model Source: Data vs. Theory
Introducing Bayesian Networks as a Research Framework
- Under the Hood: The Simple Math of Bayesian Networks
- Key Advantages of Bayesian Networks as a Modeling Framework
Seminar Program Part II (Approx. 180 minutes)
Bayesian Networks in Practice
- Knowledge Encoding & Probabilistic Inference
o Introductory Example: Prosecutor’s Fallacy
o Reinventing the Delphi Method with the Bayesia Expert Knowledge Elicitation Environment (BEKEE)
- Policy Development Under Extreme Uncertainty: Antibiotic and Anti-Malarial Prescription Guidelines in Sub-Saharan Africa
- Knowledge Discovery for Classification/Prediction
o Optimizing the Resources Required for the Diagnosis of Coronary Artery Disease
o Leukemia Classification with Microarray Analysis
- Knowledge Discovery for Human Interpretation
o Example t.b.d.
o 2D/3D/VR Visualization of Network Structures
- Knowledge Encoding + Knowledge Discovery for Causal Inference
o Simpson’s Paradox Rears its Ugly Head
o Example t.b.d.
Examples similar to the ones in this seminar can be found in our book, “Bayesian Networks & BayesiaLab: A Practical Introduction for Researchers,” which can be downloaded free of charge (https://www.bayesia.com/book)
All slides, networks, and datasets will be made available for download after the event.
PRESENTATION
Modern Machine Learning: Probabilistic Modeling and Functional Prediction
Tom Dietterich, Distinguished Professor Emeritus, Oregon State University, School of Electrical Engineering and Computer Science
Machine learning pursues two main paradigms for data analysis: probabilistic programming and function fitting. Probabilistic programming methods provide rich languages for defining probabilistic models and efficient algorithms for fitting those models to data. Function fitting algorithms seek to fit a highly-accurate prediction function drawn from a highly-flexible (non-parametric) class of functions. Building on the Bayesia tutorial from the previous afternoon, this talk will illustrate probabilistic programming by discussing multilevel modeling in the probabilistic programming language Stan. Then the talk will describe the random forest method and present recent techniques for making inferences based on these versatile models. Finally, the talk will discuss deep neural networks and their potential application to problems of environmental health science. These novel methods are best applied to problems of converting unstructured data (images, text) to structured data for subsequent statistical analysis. They also have potential to revolutionize medical imaging.
PRESENTATION
Systematic Reviews, Machine Learning and the Liberation of Knowledge from Information in Environmental Health Research
Paul Whaley, Evidence Based Toxicology Collaboration Research Fellow, Lancaster Environment Centre, Lancaster University, United Kingdom
Research synthesis methodologists have more-or-less won the argument that systematic methods represent the gold standard for reviewing and mapping scientific evidence in the service of chemicals policy and risk management decision-making. The problem is, being systematic requires a lot of resources: even to answer very focused questions which only include a few dozen primary manuscripts, it can take a typical research team 12-18 months to complete a systematic review – and there are thousands of chemicals which need reviewing over hundreds of health end-points, with results refreshed every couple of years or so. This is already more data than humans can process, and new data is piling up at an exponential rate.
Maintaining the highest standards for evidence reviews therefore requires that we get machines to read scientific documents on our behalf. This in turn requires data infrastructures which can store, and allow nuanced querying of, the very large amount of data points and complex relationships between them which are expressed in scientific documents. In this context, this presentation will give an overview of at least one way in which people who are not computer scientists can contribute to the AI revolution in chemical risk research, why we should get excited about graph databases, and what the likely milestones are between where we are now and the bold new world which is still, let’s be honest, a little way off just yet.
PRESENTATION
Environmental Health … in Context
Samuel Adams, Senior Artificial Intelligence Researcher, RTI International
Why isn’t the Holy Grail of data science and AI standing before us right now filled to overflowing with all the knowledge and insight we “just know” is there? One of the biggest reasons is the lack of context. The environment is a vast, intermeshed metasystem of metasystems where nearly everything is connected to, and interacts with, everything else, at least at some distance across a network of intermediary systems. New tools like Deep Learning hold great promise, but without putting both the inputs AND the outputs into a greater integrated context, the ultimate value delivered by our collective efforts will still be minimal. This talk will discuss both the opportunities for developing and maintaining large scale contextual knowledge graphs as well as approaches to overcoming the technical challenges along the way.
PRESENTATION
Case Studies Addressing the Following Questions
What data sets are currently available and what kind of data sets do people need to start collecting?
How to bring these techniques and methods to scientists and decision makers. How to train those people working in risk assessment. How to have them understand that this is a different paradigm.
What are the curriculum adjustments?
Harnessing Machine Learning to Predict Toxicities
Nicole Kleinstreuer, Deputy Director, National Institute of Environmental Health Sciences, NTP Interagency Center for the Evaluation of Alternative Toxicological Methods (NICEATM)
Traditional toxicology has relied upon testing chemicals in animals, methods which are costly, time-consuming, low-throughput, and often provide little insight into the mechanisms by which chemicals might affect human health and disease pathways. Toxicology testing in the 21st century, as demonstrated by the federal Tox21 research consortium, provides new methods to rapidly and reliably test tens of thousands of chemicals in a vast array of cellular, molecular, and genetic targets. When combined with large, well-curated reference datasets for substances with demonstrated toxic potential, machine learning approaches such as support vector machines, random forests, and deep learning can be used in supervised and unsupervised ways to predict complex toxicities and find associations between chemical structure and mechanisms. Such models demonstrate the power of artificial intelligence to inform regulatory decision making and help industry design safer, more sustainable products.
PRESENTATION
Using Bayesian networks to discover relations between genes, environment, and disease
Mark Borsuk, Pratt School of Engineering, Duke University
We review the applicability of Bayesian networks (BNs) for discovering relations between genes, environment, and disease. By translating probabilistic dependencies among variables into graphical models and vice versa, BNs provide a comprehensible and modular framework for representing complex systems. We first describe the Bayesian network approach and its applicability to understanding the genetic and environmental basis of disease. We then describe a variety of algorithms for learning the structure of a network from observational data. Because of their relevance to real-world applications, the topics of missing data and causal interpretation are emphasized. The BN approach is then exemplified through application to data from a population-based study of bladder cancer in New Hampshire. We find that allowing for network structures that depart from a strict causal interpretation enhances our ability to discover complex associations including gene-gene (epistasis) and gene-environment interactions. While BNs are already powerful tools for the genetic dissection of disease and generation of prognostic models, there remain some conceptual and computational challenges. These include the proper handling of continuous variables and unmeasured factors, the explicit incorporation of prior knowledge, and the evaluation and communication of the robustness of substantive conclusions to alternative assumptions and data manifestations.
PRESENTATION
Machine Learning in Dose-Response Assessment: Translating Science to Decisions
Jackie MacDonald Gibson, Associate Professor, UNC Chapel Hill, Department of Environmental Sciences and Engineering; RTI University Scholar, 2017-2018
In the United States, decisions about whether to implement new drinking water regulations are based on quantitative risk assessments of the potential health benefits of these regulations. The methods used for these risk assessments do not yet incorporate major advances in artificial intelligence that could improve risk predictions. To demonstrate the potential advantages of adopting artificial intelligence methods for United States environmental regulatory risk assessment, this talk will present a case study predicting the benefits of decreasing arsenic exposure in drinking water using a machine-learned Bayesian belief network model. The model was learned from a data set of 1,050 individuals from an arsenic-endemic region of Chihuahua, Mexico. The model integrates arsenic exposure data with biomarkers of arsenic metabolism and demographic characteristics to quantify the probability of diabetes for different exposure levels and population subgroups. The predictive ability of the Bayesian network model will be compared to that of a reference dose model and of a model estimated with Benchmark Dose Software, which are the prevailing approaches in current United States regulatory risk assessment practice. Implications for policymaking will be discussed.
PRESENTATION
Bayesian Inference for Substance and Chemical Toxicity (BISCT)
Lyle D. Burgoon, Leader, US Army Engineer Research and Development Center, Bioinformatics and Computational Toxicology Group
Lyle Burgoon leads research on Artificial Intelligence to Drive the Military Environment. The focus of his work is the creation of Artificial Intelligence that can augment human decision-making in primarily military and humanitarian environments. He is an expert in AI sensor fusion (sensors including Internet of Things sensors, ground-based sensors, space-based sensors, laboratory-based sensors) to augment human decision-making in difficult national security settings. Environmental public health challenges that he works on include predicting the impacts of the environment and environmental changes on military intelligence and warfighter readiness, AI sensor fusion to understand urban warfare environments and health impacts, to technologies for food security and logistics, to forecasting the potential secondary and tertiary impacts on human health as a result of warfare, and the automated global identification of human infrastructure networks anywhere in the world. His recent work also includes forecasting potential toxicity of military materials based on structural information, and the use of Bayesian Networks to fuse data from laboratory assays to predict potential toxicity.
PRESENTATION
Rare Diseases and AI Analysis with Potential Use of the Environmental Genome
Dr. Michael Kowolenko, CEO, NoviSystems
Dr. Michael Overcash, Executive Director, Environmental Genome Initiative
The ability to develop analytical infrastructure that allows for the fusion of different data sets can result in applications that allow users to collect information and draw conclusions in a less cumbersome and confusing fashion. This application of “backend” compute techniques such as machine learning and natural language processing allows complex relationships to be explored as a method to convey complex relationships to individuals with Rare Diseases. In a related effort, the individuals with an occurrence of rare diseases are grouped into familial or nonfamilial categories. The latter can be looked at with several analytics built on the Environmental Genome which estimates specific chemical emissions from manufacturing plants, transportation areas, the energy grid locations, and agriculture. The future challenge is how to link these chemical sources to the exposure zones of those with nonfamilial rare diseases. This is the first step in a much more transparent, but complicated approach to environmental pollutants and the use of artificial intelligence.
PRESENTATION
