Skip to main content
Back to top

While you may not see it, machine learning applications are rapidly becoming a regular part of our daily lives, changing everything from the way we practice medicine to how we browse for movies to watch on a streaming service. And yet even as the technology continues to provide new solutions and methods for organizing information, mathematicians and computer scientists are also grappling with one of machine learning’s biggest limitations.

“With machine learning, a lot of people regard it as a black box technique. You get an input and then the machine learns something, then produces an output based on that,” Dr. Barnabas Bede, program director for the DigiPen BS in Computer Science in Machine Learning, says. What happens in between is often a mystery, but one that Bede and others are actively trying to solve.

“How can we make machine learning more interpretable? How can we understand what the machines are learning?” he says.

That question, Bede says, is also the focus of an upcoming conference he’s helping to organize, the annual meeting of the North American Fuzzy Information Processing Society, or NAFIPS. The conference, which is sponsored by DigiPen Institute of Technology, was initially going to be hosted on campus in Redmond. However, due to the COVID crisis, this year’s event will instead happen online, August 20-22.

“What we plan to do is to have multiple Zoom meetings in parallel,” Bede says. “And then we will have a big Zoom meeting going on for social gatherings.”

The NAFIPS organization, for which Bede serves as a board member, aims to further mathematical research in the areas of fuzzy logic and fuzzy sets. A concept first described in 1965, fuzzy logic offers a method by which to model uncertainty for things that fall outside the scope of what can be measured in terms of probability or binary logic. Fuzzy systems are commonly used in the design and development of control systems — such as for autonomous vehicles, heating and cooling systems, and everyday household appliances to name a few.

“A fuzzy system is able to reproduce a rule-based reasoning. Like if velocity is low, then acceleration should be high, things like that,” Bede says. “So such rules can be more easily interpreted.”

One of the main themes of this year’s NAFIPS conference, he says, is how fuzzy systems can help with the inscrutability problem of traditional machine learning paradigms.

Bede, for example, will be presenting on some of his newest research into a type of fuzzy system known as the Takagi-Sugeno fuzzy system, which when used in conjunction with a neural network can create an interpretable neural network.

“I was able to find the connection between one type of fuzzy system and artificial neural networks in general,” Bede says. “That basically makes the neural network that we were working on interpretable, so that’s the good news about it. [We’re] trying to understand what machine learning is doing behind the curtain.”

Bede highlighted another panel titled “Fuzzy Logic-based eXplainable AI – An Adopter’s Perspective,” featuring Dr. Kelly Cohen, endowed chair and interim head of the Department of Aerospace Engineering & Engineering Mechanics at the University of Cincinnati, and Nicholas Ernest, chief architect of Thales USA, Inc.

“They are one of the adopters of this new explainable AI paradigm in the industry. And they will explain what they are working on,” Bede says.

Other scheduled talks include “Deep Learning (Partly) Demystified” by Dr. Vladik Kreinovich, “Privacy-preserving Machine Learning” by Dr. Martine De Cock, and “Bias and Algorithmic Discrimination in Machine Learning” by Dr. Golnoosh Farnadi, among others.

MS in Computer Science student Dhrumil Shukla will also be giving a talk, alongside Bede and two other DigiPen faculty — Dr. Natalia Solorzano, chair of the Department of Physics, and Dr. Jeremy Thomas, program director for the BS in Computer Engineering. That talk will discuss the results of Shukla’s recent thesis, which dealt with the clustering of tropical cyclones based on image processing, using multi-channel satellite imagery, and fuzzy class studying.

“Basically, we are trying to identify patterns in the tropical storm that will help us predict its behavior in the future,” Bede says.

Anyone who would like to join the conference as an online participant can register on the NAFIPS website by Saturday, August 15. The cost is $20.