Raginsky receives grant focused on information processing in belief space

8/23/2016 August Schiess, CSL

CSL Assistant Professor Maxim Raginsky will explore more computationally efficient methods for better forecasting.

Written by August Schiess, CSL

CSL Assistant Professor Maxim Raginsky has received a $75,000 grant from the National Science Foundation (NSF) Center for Science of Information to investigate methods that process information in the space of beliefs, with the goal of enabling computers to make more accurate predictions with “soft” data while consuming less energy.

Max Raginsky
Max Raginsky
Max Raginsky
Many types of decisions are stated in the form of beliefs: weather forecasts, financial projections, medical diagnostics, and more. Raginsky and his team plan to find techniques that make the process faster and perform with better accuracy, and with fewer resources (such as time, energy, or memory).

“Each time two systems communicate data and exchange their beliefs or forecasts, it uses energy,” said Raginsky, an assistant professor of electrical and computer engineering. “We want to make this process faster and easier. If you throw more energy into a system, it’s going to be better, but energy costs something. The goal is to get the same information, but with minimal resources.”

The grant, titled “Towards a Science of Information Processing in Belief Space,” aims to specifically examine the best ways for systems to convert data into beliefs and forecasts.

“A lot of the time, when some data are observed, what we get to see is not a concrete prediction, but a numerical assessment of the chances of various outcomes—that’s a probability distribution or a belief,” said Raginsky. “For example, when the weather forecast says there is a 30 percent chance of rain today, that’s a probability distribution—a belief about what’s going to happen. We want to find the best ways to navigate that belief space, to go from observations to accurate beliefs.”

In order to transform data into predictions, it is averaged and exchanged between many networks—making it more difficult to determine accurately. Raginsky aims to speed up the process by using smarter ways to identify and extract the most relevant data.

“Once you start talking about things like forecasting—exchanging beliefs or soft information—the transformations that update beliefs based on new observations are nonlinear. If we represent the observations as bits, then what happens to those bits can be very complicated,” said Raginsky. “The goal of this grant is to determine what is possible and what is impossible in systems that manipulate and generate beliefs, and then apply it to concrete problems.”

Raginsky is working with ECE PhD student Ehsan Shafieepoorfard, and they have started by investigating the types and amount of information needed to make a given decision.

“A lot of decisions you make are beliefs about how things are or about what you should do next. Nobody really has all the relevant information that you need—it is often not possible to take in all information,” said Raginsky. “So we’re after quantifying the minimum amount of information needed to generate accurate forecasts or beliefs.”

Share this story

This story was published August 23, 2016.