Artificial intelligence and machine learning have emerged as indispensable tools across numerous scientific disciplines, yet researchers continue to explore the boundaries and challenges of integrating these advanced technologies into their investigations. Undergraduate researcher Sonia Reilly dedicated her summer to demystifying the complex inner workings of neural networks, aiming to unravel how information traverses these sophisticated systems through the prestigious Undergraduate Research Opportunities Program (UROP). Her innovative project, focused on enhancing machine learning applications for observing natural phenomena, was guided by Sai Ravela from the Department of Earth, Atmospheric and Planetary Sciences (EAPS). As a student pursuing Course 18C (Mathematics with Computer Science), Reilly possesses a distinctive skill set perfectly suited for exploring these interdisciplinary connections.
"Deep learning has witnessed unprecedented adoption across research domains in recent years, yet the fundamental mathematics explaining its remarkable effectiveness remains poorly understood," Reilly explains. "Unlocking these mysteries will empower scientists to design next-generation learning machines with superior capabilities." Her research methodology involves meticulous analysis of algorithm evolution, tracing their progression toward final conclusions, with the ultimate objective of uncovering insights about information flow patterns, identifying computational bottlenecks, and optimizing neural network performance.
"Our objective isn't merely to accumulate vast quantities of data—we're striving to transform big data into intelligent, actionable insights," Ravela emphasizes regarding the future direction of machine learning in environmental science. "The vision encompasses developing intelligent sensing agents that collect environmental data efficiently, operating with knowledge-driven precision to gather precisely the information needed for meaningful analysis and decision-making."
For Ravela, who leads the pioneering Earth Signals and Systems Group (ESSG), advancing machine learning capabilities translates to more accurate early-warning systems for potential natural disasters. His research team concentrates on understanding Earth's complex systemic behavior, with particular emphasis on climate patterns and natural hazards. By observing natural phenomena, they develop sophisticated predictive models for dynamic environmental processes including hurricanes, cloud formations, volcanic activity, seismic events, glacier movements, and wildlife conservation strategies, while simultaneously pushing the boundaries of engineering and computational learning.
"Across all our research initiatives, collecting comprehensive spatiotemporal data remains practically impossible," Ravela notes. "We've demonstrated that employing a systems analytic approach to actively mine environmental information yields promising results." Ravela recently presented his team's latest findings—including Reilly's significant contributions—at the Association of Computing Machinery's Special Interest Group on Knowledge Discovery and Data Mining (SIGKDD 2019) conference. He also teaches an innovative "infinite course" series offered during spring and fall semesters, providing comprehensive foundations in machine learning for natural systems science, with materials accessible to online learners worldwide.
According to Ravela, should Reilly succeed in advancing the mathematical foundations for computational learning models, she would join the ranks of "early pioneers in explainable AI"—a groundbreaking achievement that could launch an extraordinary career trajectory.
This prospect aligns perfectly with Reilly's aspirations to pursue a PhD in mathematics after graduating from MIT, while continuing to contribute to research with global positive impact. She's maximizing her research opportunities during her final two undergraduate years at MIT, building upon this summer's transformative experience.
Although this marked Reilly's first UROP experience, it wasn't her initial foray into interdisciplinary research combining mathematics, computer science, and Earth science. Previously, at the Johns Hopkins University Applied Physics Laboratory, Reilly contributed to developing advanced signal processing techniques and software designed to extract valuable climate change information from low-quality satellite data.
"I've always envisioned myself working within interdisciplinary research environments where I could apply mathematical expertise to support scientific and engineering breakthroughs," Reilly reflects on her experience within EAPS. "Witnessing such an ecosystem firsthand has been incredibly inspiring, offering a glimpse into my potential future career path."
Ravela emphasizes that ESSG values the mutually beneficial relationship with UROP students. "In my experience, UROP students can outperform graduate students and postdocs when we craft appropriately challenging questions that match their capabilities. Given the right parameters, they demonstrate remarkable speed and ingenuity," he observes. He regards the UROP program as exceptionally valuable, believing all students could benefit from such opportunities to explore diverse fields, engage in interdisciplinary research, and learn to translate theoretical knowledge into practical applications.
For Reilly, research builds upon the structured foundation established through MIT coursework, which provides controlled, predictable learning environments. "Research, however, follows anything but a linear trajectory," she acknowledges. She's leveraged her mathematical and computational background during her UROP experience while simultaneously mastering how to connect and apply these principles to unfamiliar domains and explore concepts typically beyond undergraduate curricula. "Every research step seems to require understanding entirely new mathematical fields, making it challenging to determine where to begin. While I sometimes feel disoriented, I'm simultaneously acquiring knowledge at an extraordinary rate," she admits.
