Soheil Feizi Receives Amazon Research Award

The prestigious award has been granted for his remarkable work on mitigating spurious correlations in deep learning.  
Descriptive image for Soheil Feizi Receives Amazon Research Award

 

Soheil Feizi, an assistant professor of computer science at the University of Maryland, received an Amazon Research Award to continue his work mitigating spurious correlations in deep learning. These correlations occur when deep learning models mistakenly identify irrelevant or coincidental relationships in the data, leading to inaccurate results and potential biases.

The award program aims to foster innovation and support cutting-edge research in machine learning, computer vision, and natural language processing. 

Feizi is one of 79 award recipients from 54 universities across 14 countries. The award recognizes Feizi's dedication to advancing the field of artificial intelligence (AI) and addressing challenges in deep learning algorithms. 

“I have seen several problems in deep learning algorithms ranging from reliability, interpretability, generalization, and fairness and bias aspects,” said Feizi, who also holds a joint appointment in the University of Maryland Institute for Advanced Computer Studies. “But there is no one-size-fits-all solution for all these problems. So, we need to dig a little deeper and understand the root cause of the problems and, equally as important, understand how they may all be connected.”   

By delving into the intricate mechanisms behind spurious correlations, Feizi aims to develop innovative approaches that enable deep learning algorithms to better differentiate between meaningful correlations and chance associations, thus improving the accuracy and fairness of AI systems.

“I have observed that AI heavily relies on spurious features,” Feizi shared. “These models are very good correlation extractors, so if you have some correlation in your data that doesn’t have any links, it will not be able to differentiate between the two, which may cause AI to give unclear answers that don’t make sense to humans. And this is becoming a theme. People ask questions and get answers that are farfetched and off-topic. With my research, I want to know why this is happening and how we could do something about it moving forward.”  

Feizi's research has significant implications in numerous fields, including healthcare and finance. By addressing spurious correlations, his work has the potential to revolutionize how deep learning models are developed and applied, leading to more reliable and trustworthy AI-driven solutions.

“Today, we have AI models in many places like speech processing, language models and vision, and wherever we have these complex workings, we also will have the issues of spurious correlations,” Feizi shared. “This is one of the fundamental issues that show unreliability in deep learning. I look forward to researching more on this to help bridge the gaps between AI and dependability.”    

Feizi sees the award as a springboard to take on even more significant challenges and is delighted to be a part of something that can bring meaningful change to the field.

“I am deeply honored to receive the Amazon Research Award, as it underscores the significance of combating spurious correlations in deep learning,” Feizi said. “This grant will enable me to explore new avenues to enhance the integrity and interpretability of AI algorithms, benefiting various applications across many industries.”

The recipients of the Amazon Research Awards are granted numerous advantages to enhance their research endeavors, including access to over 300 public datasets and Amazon Web Services (AWS) Promotional Credits, which provide them with privileged access to a diverse range of AWS services and tools.

—Story by Samuel Malede Zewdu, CS Communications 

The Department welcomes comments, suggestions and corrections.  Send email to editor [-at-] cs [dot] umd [dot] edu.