Deep neural networks demonstrate exceptional capability in identifying complex patterns within massive datasets that exceed human cognitive processing capacity. This remarkable proficiency has established deep learning as an essential tool for virtually all data professionals. This year, the MIT Quest for Intelligence and the MIT-IBM Watson AI Lab provided funding for 17 talented undergraduates to collaborate with faculty members on year-long research initiatives through MIT's prestigious Advanced Undergraduate Research Opportunities Program (SuperUROP).
Participants had the unique opportunity to investigate practical AI applications in climate science, finance, cybersecurity, and natural language processing, among diverse domains. Meanwhile, faculty members benefited from collaborating with students from different academic departments, an experience they enthusiastically endorse. "Adeline exemplifies the tremendous value of the UROP program," notes Raffaele Ferrari, a professor in MIT's Department of Earth and Planetary Sciences, referring to his research advisee. "Without UROP, an oceanography professor might never have had the chance to collaborate with a computer science student."
Below we highlight four groundbreaking SuperUROP projects from the past academic year.
Advanced Algorithm Development for Cloud Computing Optimization
The transition from traditional desktop computing to distributed data centers in the "cloud" has created significant operational challenges for companies providing computing services. Facing continuous fluctuations in service requests and cancellations, their profitability largely depends on efficiently matching computing resources with client needs.
Approximation algorithms are employed to accomplish this complex optimization task. Among all possible approaches for allocating machines to customers based on pricing and other factors, these algorithms identify schedules that achieve near-optimal profit margins. Over the past year, junior Spencer Compton collaborated on a virtual whiteboard with MIT Professor Ronitt Rubinfeld and postdoc Slobodan Mitrović to develop a more efficient scheduling methodology.
"We didn't write any code," Compton explains. "We developed mathematical proofs and applied theoretical concepts to discover a more efficient approach to solving this optimization challenge. The same principles that enhance cloud-computing scheduling can be applied to assigning flight crews to aircraft and numerous other logistical problems."
In a pre-print paper published on arXiv, Compton and his research team demonstrate how to accelerate an approximation algorithm under dynamic conditions. They also present techniques for locating specific machines assigned to individual customers without processing the entire schedule.
"Identifying the core focus of the project presented a significant challenge," Compton admits. "There's extensive existing literature, and many researchers have examined related problems. It was fascinating to review previous work and brainstorm areas where we could make meaningful contributions."
Predicting Ocean Heat and Carbon Absorption Capacity Using AI
Earth's oceans play a crucial role in climate regulation by absorbing excess heat and carbon dioxide from the atmosphere. However, as ocean temperatures rise, scientists question whether they will continue to absorb carbon at current rates. A reduced uptake could lead to accelerated warming beyond current climate model projections. This represents one of the critical questions confronting climate scientists as they strive to enhance their predictive capabilities.
The primary obstacle in addressing this question is the problem's inherent complexity: current global climate models lack the computational power to provide high-resolution insights into the dynamics affecting key variables such as sea-surface temperatures. To compensate for this limitation, researchers are developing surrogate models to approximate these missing dynamics without explicitly calculating them.
In collaboration with MIT Professor Raffaele Ferrari and research scientist Andre Souza, MIT junior Adeline Hillier is investigating how deep learning models for ocean climate prediction can enhance or potentially replace physical models of the ocean's uppermost layer, which governs heat and carbon absorption rates. "If the model maintains minimal computational requirements while performing effectively under diverse real-world physical conditions, it could be integrated into global climate models and potentially improve climate projections," Hillier explains.
Throughout the project, Hillier acquired programming skills in Julia and received an intensive introduction to fluid dynamics. "You're attempting to model the effects of turbulent dynamics in the ocean," she says. "Understanding the underlying processes and physics significantly enhances the modeling process."
Developing Next-Generation Efficient Deep Learning Architectures
There exist countless approaches to designing a deep learning model for a specific task. Automating this design process promises to streamline options and make these tools more widely accessible. However, identifying the optimal architecture is remarkably complex. Most automated searches select models that maximize validation accuracy without considering the underlying data structure, which might suggest simpler, more robust solutions. Consequently, more reliable or data-efficient architectures are often overlooked.
"Rather than focusing exclusively on model accuracy, we should prioritize understanding data structure," emphasizes MIT senior Kristian Georgiev. Working with MIT Professor Asu Ozdaglar and graduate student Alireza Fallah, Georgiev is exploring methods to automatically analyze data to identify the most suitable architecture for its specific constraints. "When you select your architecture based on data characteristics, you're more likely to achieve a robust, high-quality solution from a learning theory perspective," he explains.
"The initial exploratory phase presented the greatest challenge," Georgiev recalls. To identify a compelling research question, he reviewed papers spanning topics from automated machine learning to representation theory. However, the effort proved worthwhile, enabling him to work at the intersection of optimization and generalization. "Significant progress in machine learning requires integration of both these fields," he notes.
Decoding Human Face Recognition Superiority Through AI Research
Face recognition comes naturally to humans. Identifying familiar faces in blurred or distorted photographs poses little difficulty. However, we don't fully comprehend the mechanisms behind this ability or how to replicate this capability in machines. To identify the principles crucial to face recognition, researchers have presented progressively degraded headshots to human subjects to determine when recognition begins to fail. They are now conducting similar experiments with computers to determine whether deeper insights can be obtained.
In a project with MIT Professor Pawan Sinha and the MIT Quest for Intelligence, junior Ashika Verma applied various filters to a dataset of celebrity photographs. She blurred facial features, applied distortions, and modified colors to evaluate whether a face-recognition model could identify photographs of the same individual. She discovered that the model performed best with either natural color or grayscale images, consistent with human studies. When color filters were applied, accuracy decreased, though not as dramatically as with human subjects—a discrepancy Verma plans to investigate further.
This research contributes to a broader initiative to understand human superiority in face recognition and how machine vision might be enhanced accordingly. The project also connects with Project Prakash, a nonprofit organization in India that treats blind children and monitors their recovery to advance understanding of the visual system and brain plasticity. "Human experiments require significantly more time and resources than computational studies," explains Verma's advisor, Kyle Keane, a researcher with MIT Quest. "We're striving to make AI as human-like as possible to conduct numerous computational experiments that identify the most promising approaches for subsequent human studies."
"Preparing the degraded images for experiments and processing them through deep networks presented technical challenges," Verma acknowledges. "The process is very slow—you work for 20 minutes at a time and then wait." However, the laboratory experience with expert guidance made the effort worthwhile, she says. "It was fascinating to gain exposure to neuroscience."
SuperUROP projects received partial funding from the MIT-IBM Watson AI Lab, MIT Quest Corporate, and from Eric Schmidt, technical advisor to Alphabet Inc., and his wife, Wendy.
