New tool brings benefits of AI programming to decision making under uncertainty

One reason deep learning has exploded over the past decade is the availability of programming languages ​​that can automate the math—college-grade calculus—needed to train each new model. Neural networks are trained by adjusting their parameters to try to maximize a score that can be quickly computed for the training data. The equations used to adjust the parameters in each tuning step were carefully derived by hand. Deep learning platforms use a method called automatic discrimination to automatically compute the adjustments. This allows researchers to quickly explore a large space of patterns and find ones that actually work without knowing basic math.

But problems like climate modelor financial plan, where the underlying scenario is uncertain? For these problems, calculation alone is not enough—you also need probability theory. “Score” is no longer just a defined function of parameters. Instead, it is determined by a random model that makes random choices to model the unknowns. If you try to use deep learning platforms for these problems, they can easily give the wrong answer. To solve this problem, MIT researchers developed ADEV, which extends automatic differences to handle models that make random choices. This brings the benefits of AI programming to a wider variety of problems, allowing for rapid experimentation with models that can infer about uncertain situations.

Lead Author and MIT electrical engineering and PhD in computer science. Student Alex Lew says he hopes people will be less wary of using probabilistic models because there is now a tool to automatically distinguish them. “The need to manually draw unbiased, low-variance slope estimators may lead to the perception that probabilistic models are more complex or complex to work with than deterministic models. But probability is an incredibly useful tool for modeling the world.I hope that by providing a framework for building these estimators automatically, ADEV will make testing probabilistic models becomes more compelling, possibly enabling new discoveries and advancements in AI and beyond.”

Sasa Misailovic, an associate professor at the University of Illinois at Urbana-Champaign, who was not involved in the study, added: “As probabilistic programming paradigms are emerging to solve various problems in science and engineering, the question is how can we make efficient software implementations built on solid mathematical principles ADEV presents such a foundation for probabilistic inference components and modulus with derivatives ADEV brings the benefits of probabilistic programming—automatic mathematics and more scalable inference algorithms—to a much broader range of problems whose goals are not achieved only to deduce what might be true but also to decide on the next course of action.”

In addition to climate modeling and financial modeling, ADEV can also be used for research activities—for example, simulating a customer queue to a call center to minimize expected wait times, by simulating the waiting process and evaluating the quality of the results—or to tweak the algorithm that -Boots used to capture physical objects. Co-author Mathieu Huot said he was pleased to see ADEV “used as the design space for new low-variance estimators, a key challenge in probability computation.”

The study, which was awarded the SIGPLAN Outstanding Paper at POPL 2023, was co-authored by Vikash Mansighka, head of MIT’s Probabilistic Computing Project in the Department of Cognitive and Brain Sciences as well as the Science Laboratory. Computing and Artificial Intelligence, and helping lead the MIT Quest for Intelligence, as well as Mathieu Huot and Sam Staton, both at the University of Oxford. Huot adds, “ADEV provides a unified framework for arguing the common problem of slope estimation in an unbiased, clean, elegant and structured way.”

“Many of our most controversial decisions—from climate policy to tax codes—focusing on decision-making under uncertainty. ADEV makes it easier to test new ways to solve these problems, by automating some of the most difficult math operations,” says Mansinghka. probabilistic program, we have new, automated ways to adjust parameters accordingly. try to produce the results we want and avoid the results we don’t want.”

This story is republished with permission from MIT News (web.mit.edu/newsoffice/), a popular website that covers MIT research, innovation, and teaching.