Artificial intelligence systems have provided scientists with unprecedented capabilities for future forecasting and predictive analytics. However, these neural network technologies face significant challenges due to their enormous appetite for data processing resources and computational power. At MIT, researchers require approximately five times more computing infrastructure than currently available. To address this computational gap, technology industry partners have stepped forward with critical support. A state-of-the-art $11.6 million supercomputer, recently contributed by IBM, became operational this autumn, while both IBM and Google have provided valuable cloud computing credits to MIT Quest for Intelligence for distribution throughout the campus community. Below, we highlight four innovative research projects made possible through these industry partnerships.
Revolutionizing Neural Network Efficiency
For deep learning models to accurately identify objects like cats in photographs, they traditionally require processing millions of images before their artificial neurons can effectively recognize patterns. This training methodology demands substantial computational resources and carries significant environmental consequences, as recent studies measuring AI's carbon footprint have demonstrated.
However, groundbreaking research from MIT suggests a more efficient approach exists. Scientists have discovered that models merely a fraction of conventional sizes could achieve comparable results. "When training extensive networks, there exists a compact alternative capable of performing identical tasks," explains Jonathan Frankle, a graduate student in MIT's Department of Electrical Engineering and Computer Science (EECS).
Collaborating with EECS Professor Michael Carbin, Frankle demonstrated that neural networks might function effectively with merely one-tenth the typical connections when the optimal subnetwork is identified initially. Standard practice involves pruning neural networks after training completion, removing unnecessary connections. This led Frankle to question: why not begin with the smaller model from the start?
After conducting preliminary experiments with a basic two-neuron network on his personal laptop, Frankle observed promising results and progressed to larger image datasets including MNIST and CIFAR-10, utilizing available GPUs whenever possible. Finally, leveraging IBM Cloud resources, he obtained sufficient computational power to train a genuine ResNet model. "All my previous work represented merely experimental trials," he notes. "I could finally execute numerous configuration variations to validate the claims presented in our research paper."
Frankle shared these insights from Facebook's headquarters, where he spent the summer exploring concepts introduced in his Lottery Ticket Hypothesis paper, one of only two selected for a best paper award at this year's International Conference on Learning Representations. The practical applications of this research extend beyond image classification to encompass reinforcement learning and natural language processing models. Researchers at Facebook AI Research, Princeton University, and Uber have already published follow-up studies building upon this foundation.
"What fascinates me most about neural networks is that we haven't even established their fundamental principles yet," reflects Frankle, who transitioned from studying cryptography and technology policy to artificial intelligence. "We lack comprehensive understanding of how these systems learn, where they excel, and where they fall short. We're essentially where physics was a millennium before Newton."
Advanced AI Systems for Detecting Misinformation
Social media platforms including Facebook and Twitter have transformed how we access news information. However, legitimate journalism frequently becomes overshadowed by misleading or completely fabricated content circulating online. The recent controversy surrounding a manipulated video of U.S. House Speaker Nancy Pelosi, altered to make her appear intoxicated, represents just one example of how misinformation threatens democratic discourse.
"Virtually anything can be published online today, and some portion of the audience will accept it as fact," observes Moin Nadeem, a senior and EECS major at MIT.
If technology contributed to creating this challenge, it can also provide solutions. This reasoning motivated Nadeem to select a superUROP project focused on developing an automated system to combat fake and misleading news. Working in the laboratory of James Glass, a researcher at MIT's Computer Science and Artificial Intelligence Laboratory, and supervised by Mitra Mohtarami, Nadeem contributed to training a language model designed to fact-check claims by searching through Wikipedia and three categories of news sources evaluated by journalists as high-quality, mixed-quality, or low-quality.
To verify a specific claim, the model evaluates the degree of consensus among sources, with higher agreement scores indicating the claim's likely veracity. A high disagreement score for a statement such as "ISIS infiltrates the United States" strongly suggests fake news. Nadeem acknowledges one limitation of this approach: the model doesn't establish objective truth so much as reflect prevailing opinions.
Utilizing Google Cloud Platform resources, Nadeem conducted experiments and developed an interactive website enabling users to instantly evaluate claim accuracy. He and his research collaborators presented their findings at the North American Association of Computational Linguistics (NAACL) conference in June and continue expanding their research efforts.
"The traditional wisdom held that seeing was believing," Nadeem explains in this video about his research. "However, we're entering an era where that no longer applies. When people cannot trust their visual and auditory perceptions, we must ask: what can we trust?"
AI-Powered Climate Change Visualization
From rising sea levels to increased drought frequency, the impacts of climate change are already evident globally. Within several decades, our planet will likely become warmer, drier, and increasingly unpredictable. Brandon Leshchinskiy, a graduate student in MIT's Department of Aeronautics and Astronautics (AeroAstro), is exploring generative adversarial networks (GANs) to visualize Earth's potential future appearance.
GANs create remarkably realistic imagery by setting two neural networks in competition with each other. The first network learns the underlying patterns within a collection of images and attempts to replicate them, while the second network evaluates which images appear unrealistic and instructs the first network to improve its attempts.
Inspired by researchers who utilized GANs to visualize sea-level rise projections from street-view perspectives, Leshchinskiy investigated whether satellite imagery could similarly personalize climate forecasts. Working with his advisor, AeroAstro Professor Dava Newman, Leshchinskiy currently employs complimentary IBM Cloud credits to train paired GANs using images of the eastern U.S. coastline alongside corresponding elevation data. The objective involves visualizing how sea-level rise projections for 2050 might reshape coastal boundaries. If successful, Leshchinskiy plans to incorporate additional NASA datasets to imagine future ocean acidification scenarios and phytoplankton population changes.
"We've moved beyond the point where mitigation alone suffices," Leshchinskiy states. "Visualizing our world three decades from now can help us develop adaptation strategies for climate change."
AI-Powered Athlete Identification Through Motion Analysis
Merely observing a few athletic movements enables computer vision models to identify individual players with remarkable accuracy. This finding emerges from preliminary research conducted by a team led by Katherine Gallagher, a researcher at MIT Quest for Intelligence.
The research team trained computer vision models using video recordings from tennis matches, soccer games, and basketball competitions. They discovered that these models could recognize individual athletes within just a few frames by analyzing key body points that provide a basic skeletal outline.
The team utilized a Google Cloud API to process the video data, comparing their models' performance against those trained on Google Cloud's AI platform. "This pose information proves so distinctive that our models can identify players with accuracy nearly matching models provided with substantially more information, such as hair color and clothing details," Gallagher explains.
These findings offer valuable applications for automated player identification within sports analytics systems. Additionally, this research could establish foundations for further studies on inferring player fatigue levels to optimize substitution timing. Automated pose detection might also assist athletes in refining their techniques by enabling precise isolation of specific movements associated with expert performance, such as a golfer's powerful drive or a tennis player's winning swing.
