If the outcome is uncertain, how do you weigh the competing value? If complete information is not available, what constitutes a reasonable choice? Once confined to academic philosophy, these questions are now at the forefront and centre, as they delegate increasingly complex decisions to AI.
A new large-scale linguistic modeling framework developed by Willie Neiswanger, assistant professor of computer science at USC Viterbi Engineering School of Advanced Computing, along with students from Thomas Lord's Computer Science department at USC Viterbi, we combine classical decision theory and utility theory principles to face AI uncertainty and enhance complex decisions.
Neiswanger's research was highlighted at this year's international conference on learning expressions. He discussed how AI handles uncertainty on USC News.
What do you think about the difference between artificial intelligence and human intelligence?
Neiswanger: Today, human intelligence has many advantages compared to mechanical intelligence. However, machine intelligence has certain strengths and value compared to humans. Large-scale Language Models (LLMS) – AI systems trained with huge amounts of text that can understand and generate human-like responses can be generated at scale by rapidly ingesting and synthesizing large amounts of information from reports and other data sources, simulating many possible futures, or suggesting a wide range of predicted results. In our work, we aim to balance the strengths of LLMS with human strengths and judgment.
Why do today's leading language models of AI suffer from uncertainty?
Neiswanger: Uncertainty is a fundamental challenge in real decision-making. Current AI systems struggle to balance the process of making uncertainty, evidence, and prediction based on the likelihood of unknown outcomes and user preferences when faced with unknown variables.
Unlike human experts who can express the degree of confidence and acknowledge the limits of knowledge, LLMs usually generate responses with obvious confidence, whether they draw from established patterns or make uncertain predictions beyond the available data.
How does your research intersect with uncertainty?
Neiswanger: I focus on developing machine learning methods for decision making under uncertainty. We focus on sequential decision-making. This is a situation in which each decision affects future options and in a setting that is expensive for data to acquire, each decision affects future options. This includes applications such as black box optimization (finding the best solution if you don't know how the system works internally), experimental design (design research or testing to get the most useful information), and science and engineering decision-making tasks (materials and drug discovery, etc.).
I am also interested in large AI systems that are trained on huge datasets that act as large language models), especially large language models (how both the enhancements and benefits of these decision-making frameworks work).
How did your research address uncertainty and AI issues?
Neiswanger: We focused on improving the machine's ability to quantify uncertainty, and teaching it to measure and express how confident it should be about essentially different predictions. In particular, we can develop an uncertainty quantification approach that allows large-scale linguistic models to make decisions under incomplete information, and make predictions at measurable confidence levels that can be verified, and select actions that provide the greatest benefits to human preferences.
This process was converted into numerical possibilities based on reports, historical data and other contextual information by identifying key uncertain variables associated with decision-making and assigning language-based probability scores to the language model to various possibilities (e.g. crop yield, stock price, dates of uncertain events, predicted warehouse transport).
Are there any instant applications?
Neiswanger: In business situations, strategic planning could be improved by providing a more realistic assessment of market uncertainty and competitive dynamics. In a healthcare setting, doctors may provide diagnostic support or support in treatment planning by better explaining symptoms and uncertainties about test results. Personal decisions can help users get more informed and relevant advice from their linguistic models of everyday choices.
The ability of a system to suit human preferences is particularly valuable in the context in which computers can find mathematically “best” solutions, and can miss important human values and constraints. By explicitly modeling stakeholder preferences and incorporating them into mathematical assessments of how different outcomes are valuable to people, the framework generates recommendations that are not only technically optimal, but are actually acceptable to those who implement them.
What's next for your research?
Neiswanger: We are currently investigating ways this framework can be extended to broader real-world decisions under uncertainty tasks, such as operations research (using mathematical methods to solve complex business problems), logistics, healthcare, and other applications. One of the focuses on the future is improving human auditability. It involves developing an interface that allows users to clearly visualize why LLM makes a particular decision and why that decision is best suited.
