Ngan Le, an associate professor in the Department of Electrical Engineering and Computer Science, has been awarded the prestigious National Science Foundation Faculty Early Career Development award. This award supports the advancement of her research on developing artificial intelligence systems that are trustworthy, reliable and sustainable in real-world environments.

Le's $499,556 CAREER award will fund her work on creating a biologically inspired, trustworthy, robust and energy-efficient multimodal framework for video analytics. This research represents a critical step toward enabling AI systems that can be safely deployed in the real world. As principal investigator, Le aims to create an AI framework inspired by human perception.This framework will be capable of integrating multiple sensory inputs to deliver transparent, understandable decisions while maintaining reliability even in noisy environments or with incomplete information.

"Healthcare and robotics are likely to be the first to adopt and apply my research," Le said. "These domains require models that can seamlessly integrate heterogeneous data sources, like visual, spatial, temporal or sensor data. Moving forward, I envision these industries integrating such methods into real-time systems to enhance perception, diagnosis and decision support. This enables safer, more reliable and context-aware applications in complex and dynamic environments."

Artificial intelligence technologies are now integral to sectors such as health care, robotics, agriculture, public safety and surveillance. While AI innovation is transforming industry capabilities, it also raises critical concerns around trustworthiness, reliability and environmental sustainability. These factors pose substantial risks to the long-term scientific and societal benefits offered by AI technologies.

Le's research addresses these challenges through a novel multimodal framework that processes diverse data streams, including video, audio, text and 3D point clouds, to make decisions that are transparent and computationally efficient.

"Today's AI models often operate as black boxes. They lack explainability and accountability, struggling with noisy or incomplete data.They also consume excessive energy," Le said. "This research aims to make AI more interpretable, like the human brain, more resilient in unpredictable conditions and more sustainable during deployment."

The project is structured around three key goals:

• Trustworthy Human-Like Scene Awareness: Developing interpretable video analytics using dynamic semantic graphs that mimic human understanding of scenes, enabling systems to explain their decisions and track interactions over time. 
• Robust Multimodal Integration: Creating a collaborative expert-agent framework that fuses multiple visual and non-visual data modalities to ensure system reliability even when some inputs are missing or corrupted. 
• Efficient AI Deployment: Introducing a spectrum-preserving token merging technique, inspired by graph spectral theory, to reduce energy and computational overhead during inference. This is key to making large AI models more environmentally sustainable. 

Beyond its technological goals, Le's research includes significant educational outreach components designed to expand participation in STEM.

"My lab regularly hosts high school students who are interested in gaining hands-on experience in machine learning, but this is an informal program," Le said.

As the project grows, it will offer additional hands-on learning experiences for K-12 students, mentoring for undergraduate and graduate students, and the creation of new university courses focused on trustworthy machine learning. These efforts are intended to inspire and prepare the next generation of scientific leaders dedicated to responsible AI development.

Professor Le is also part of the UA Power Group, a group of faculty from multiple disciplines with various research focuses, all working towards elevating solutions to address global energy challenges in our electrified world.