Artificial intelligence has become increasingly integral to modern society, driving the necessity for innovative technologies that foster confidence in emerging domains, address evolving threats, and adapt to dynamic complex environments.
In a groundbreaking partnership to propel AI innovation forward, the MIT Stephen A. Schwarzman College of Computing has joined forces with Singapore's Defense Science and Technology Agency. This collaboration has resulted in funding for 13 pioneering research initiatives led by MIT scholars, focusing on three critical domains: establishing trustworthy AI systems, enhancing human cognitive capabilities in challenging settings, and democratizing AI accessibility for all. Below, we spotlight these transformative research endeavors.
"SYNTHBOX: Establishing Real-World Model Robustness and Explainability Using Synthetic Environments" spearheaded by Aleksander Madry, computer science professor. Revolutionary machine learning advancements are poised to transform tasks previously reserved exclusively for human expertise. By harnessing cutting-edge developments in realistic graphics rendering, data modeling, and inference methodologies, Madry's research team is constructing an innovative toolkit designed to accelerate the development and implementation of reliable machine learning solutions.
"Next-Generation NLP Technologies for Low-Resource Tasks" led by Regina Barzilay, Delta Electronics Professor of Electrical Engineering and Computer Science, and Tommi Jaakkola, Thomas Siebel Professor in the same field. Within natural language processing technologies, numerous global languages lack comprehensive annotation. This deficiency in direct supervision frequently yields imprecise, unreliable, and fragile outcomes. Under Barzilay and Jaakkola's guidance, the research team is creating advanced text-generation tools for controlled style transfer alongside innovative algorithms capable of identifying misinformation and suspicious online content.
"Computationally-Supported Role-playing for Social Perspective Taking" directed by D. Fox Harrell, professor of digital media and artificial intelligence. This interdisciplinary initiative merges computer science and social science methodologies to develop tools and techniques that model social phenomena within computer-supported role-playing platforms—including online gaming, augmented reality, and virtual reality—enabling users to better comprehend perspectives of individuals with diverse social identities.
"Improving Situational Awareness for Collaborative Human-Machine First Responder Teams" under the leadership of Nick Roy, aeronautics and astronautics professor. When addressing urban emergencies, achieving comprehensive situational awareness becomes paramount. Roy's team is engineering a sophisticated multi-agent system featuring autonomous aerial and ground vehicles capable of reaching emergency locations, mapping these environments to provide preliminary situation reports to first responders, and conducting searches for individuals and entities of interest.
"New Representations for Vision" guided by William Freeman, Thomas and Gerd Perkins Professor of Electrical Engineering and Computer Science, and Josh Tenenbaum, cognitive science and computation professor. A yet-unrealized AI objective involves accurately modeling the intricate shapes and textures of real-world scenes captured in images. This initiative concentrates on developing advanced neural network representations better suited to vision and graphics requirements, enabling efficient 3D world representation while capturing its complexity.
"Data-driven Optimization Under Categorical Uncertainty, and Applications to Smart City Operations" led by Alexandre Jacquillat, assistant professor of operations research and statistics. Smart city technologies offer solutions for metropolitan areas facing mounting pressures to manage congestion, reduce greenhouse gas emissions, enhance public safety, and improve healthcare delivery. Jacquillat's team is pioneering novel AI tools to manage cyber-physical infrastructure in smart cities while developing and implementing automated decision-making systems for urban operations.
"Provably Robust Reinforcement Learning" directed by Ankur Moitra, Rockwell International Career Development Associate Professor of Applied Mathematics. Building upon their innovative framework for robust supervised learning, Moitra's team is exploring more complex learning challenges, including designing resilient algorithms for reinforcement learning within Massart noise models—a research territory that remains largely unexplored.
"Audio Forensics" under the guidance of James Glass, senior research scientist. Advancements in multimedia manipulation and generation technologies—particularly speech, images, and video—have produced increasingly convincing "deepfake" content that becomes progressively difficult to distinguish from authentic material. Glass's team is developing sophisticated deep learning models designed to identify manipulated or synthetic speech content while also detecting deepfake characteristics to help analysts understand the underlying objectives of such manipulations and the resources required to create them.
"Building Dependable Autonomous Systems through Learning Certified Decisions and Control" led by Chuchu Fan, assistant professor of aeronautics and astronautics. While machine learning presents unprecedented opportunities for achieving complete autonomy, learning-based approaches in autonomous systems can fail due to suboptimal data quality, modeling errors, interactions with other agents, and complex engagements with human and computer systems in contemporary operational environments. Fan's research group is constructing a comprehensive framework comprising algorithms, theories, and software tools for learning certified planning and control systems, alongside developing firmware platforms for automatic plug-and-play quadcopter design and formation control for mixed ground and aerial vehicles.
"Online Learning and Decision-making Under Uncertainty in Complex Environments" directed by Patrick Jaillet, Dugald C. Jackson Professor of Electrical Engineering and Computer Science. Technological progress in computing, telecommunications, sensing capabilities, and information technologies offers tremendous potential for leveraging dynamic information to enhance productivity, optimize performance, and address novel complex online problems of significant practical interest. However, these opportunities present substantial methodological challenges in formulating and solving these emerging issues. Jaillet's team employs machine learning techniques to systematically integrate online optimization and learning, facilitating human decision-making under uncertain conditions.
"Analytics-Guided Communication to Counteract Filter Bubbles and Echo Chambers" led by Deb Roy, professor of media arts and sciences. Rather than expanding our worldviews as promised, social media technologies have algorithmically confined users within homogeneous information bubbles. Roy's team is developing language models and methodologies to counteract these technologies' effects, which have exacerbated socioeconomic divisions and limited exposure to diverse perspectives, restricting opportunities for users to learn from individuals with different backgrounds, beliefs, and lifestyles.
"Decentralized Learning with Diverse Data" guided by Costis Daskalakis, professor of electrical engineering and computer science; Asu Ozdaglar, MathWorks Professor and department head of EECS, and deputy dean of MIT Schwarzman College of Computing; and Russ Tedrake, Toyota Professor of Electrical Engineering and Computer Science. In numerous AI applications, combining diverse experiences and decentralized data collected by heterogeneous agents proves essential for developing superior models for prediction and decision-making across various new tasks these agents undertake. This project integrates tools from machine learning, optimization, control, statistics, statistical physics, and game theory to advance the fundamental science of federated or fleet learning—learning from decentralized agents with diverse data—utilizing robotics as an application area to provide rich, relevant data sources.
"Trustworthy, Deployable 3D Scene Perception via Neuro-symbolic Probabilistic Programs" directed by Vikash Mansinghka, principal research scientist; Joshua Tenenbaum, professor of cognitive science and computation; and Antonio Torralba, Thomas and Gerd Perkins Professor of Electrical Engineering and Computer Science. For real-world deployment, 3D scene perception systems must generalize across environments and sensor configurations while adapting to scene and environmental changes without requiring costly retraining or fine-tuning. Building upon the researchers' breakthroughs in probabilistic programming and real-time neural Monte Carlo inference for symbolic generative models, the project team is developing a domain-general approach to trustworthy, deployable 3D scene perception that addresses fundamental limitations of current state-of-the-art deep learning systems.
