MIT-Google Program for Computing Innovation
The MIT-Google Program for Computing Innovation provides funding for research projects, fellowships, cloud platform support, and research inclusion initiatives. The program will fund a mix of single- and multi-year research projects focused on global societal and sustainability solutions provided by computing, and will involve interdisciplinary teams of bilingual computing researchers, as well as support for rising researchers from historically marginalized groups. The first year of the program is focused on funding projects in the areas of computing for the health of the planet, responsible computing, and computer-aided creativity.
- “Mapping the Effects of Mountaintop Removal on Downstream Biodiversity in Appalachia”
Kerri Cahoy, Department of Aeronautics and Astronautics
This project aims to better understand the impacts of mountaintop removal on biodiversity in Appalachia by utilizing toolkits that are readily available through Google applications. It will also seek to further develop a tool that can be used to more quickly identify impacts, aiding in tracking biodiversity loss.
- “Accelerated Evolution and Discovery of Sustainable Materials Using Autonomous Experimentation”
Markus J. Buehler, Department of Civil and Environmental Engineering
The diverse array of functional materials that occur naturally is evidence of the potential for such materials to be developed from renewable resources, but creating composites from organic stocks with such properties is often a time- and labor-intensive process that limits practical applications in most cases. This project proposes an integrated autonomous platform that uses multiscale mechanics, additive manufacturing, and deep learning for predictive modeling to manufacture, characterize, and evolve composites built from a set of abundant organic feedstocks into diverse hierarchical assemblies with objective properties.
- “The Embodied Metaverse from the Ground Up”
Lawrence Sass and Nikolaos Vlavianos, Department of Architecture
Fotini Christia, Department of Political Science and Institute for Data, Systems, and Society
Munzer Dahleh, Department of Electrical Engineering and Computer Science and Institute for Data, Systems, and Society
This project sets out to create digital twins and virtual experiences of cultural spaces of importance among African American communities across the U.S. The outputs of this research would not only add a set of significant sites to the African American Metaverse, but will also enable the preservation of such sites and the creation of educational initiatives around them. The technical process of data collection and rendering of these VR spaces would involve the training of local constituents, passing down technical expertise to the participating communities.
- “Can It See What I See? In Search of Computer Vision Algorithms That Respond to Visual Illusions as Humans Do”
James DiCarlo, Department of Brain and Cognitive Sciences and MIT Quest for Intelligence
This project sets out to bring together primate vision neuroscientists and human behavioral scientists from MIT with computer vision scientists from Google to probe the computations that underlie the phenomena of visual illusions. Over the last several decades, computer vision algorithms have achieved remarkable levels of performance in a variety of tasks, including object recognition, depth perception, image segmentation, scene understanding, the artificial generation of realistic looking images and artwork. Indeed, there is now a veritable “zoo” of alternative visual processing algorithms, all of which operate on spatial patterns of luminance and some of which share at least partial alignment with primates at both the neural and behavioral levels.
- “Modular and Editable AI Image Synthesis with Generative Scene Graphs”
Frédo Durand, Department of Electrical Engineering and Computer Science and Computer Science and Artificial Intelligence Laboratory
This project aims to bring the power of scene representation data structures, such as scene graphs used in 3D software and compositing graphs used in movie special effects and photo editing software, to generative AI. It will also focus on modular and editable image synthesis as such representations and approaches could have important implications for perception tasks because of their modularity and composability, both for generating training data and for analysis itself.
- “Hybrid Physics and Data-driven Methods for Statistics of Extreme Weather Events From Climate Simulations”
Themistoklis Sapsis, Department of Mechanical Engineering
The objective of this work is the development of a systematic framework for the AI improvement of physics-driven climate models using reanalysis data, with particular emphasis on the accurate prediction of extreme weather event probabilities, such as heat waves and precipitation events. To overcome the limitations of purely data-driven or physics-based methods, this project will combine them, aiming for improved generalization properties, especially for extreme events.
- “Automated Single-Image Procedural Material Capture+A11″
Wojciech Matusik, Department of Electrical Engineering and Computer Science and Computer Science and Artificial Intelligence Laboratory
Procedural materials create photorealistic appearances for industry-grade 3D assets and are ubiquitous in producing movies, video games, digital twins, and VR/AR applications. But, designing procedural material to replicate real-world material appearances is a challenging, labor-intensive process that takes hours to days, even for professional artists, to manually connect dozens of nodes and fine-tune hundreds of node parameters. This project proposes an automated approach to capturing production-grade procedural material from single-flash photos of real object surfaces, and aims to provide essential tools that assist users throughout the procedural material authoring process and dramatically shorten their working cycle.
- “Leveraging Generative AI for Physical Object Stylization+A12”
Stefanie Mueller, Departments of Electrical Engineering and Computer Science and Mechanical Engineering, and Computer Science and Artificial Intelligence Laboratory
This project sets out to explore how to adapt generative AI models to better support physical object creation. In particular, the project will explore how generative AI can be used to create physically valid outputs for digital manufacturing, such as when creating 3D models for 3D printing. Extending generative AI models to support physical object fabrication will enable many new applications in design, art, and engineering since the resulting outputs can be physically made and function in the real world.
- “AI for Pandemic Prevention: Energy and Data Efficient Machine Learning Algorithms for Rapid Point-of-Need Diagnostic Tools for the Fight Against Climate Change Induced Global Health Threats”
Loza Tadesse, Department of Mechanical Engineering
Regions most affected by infectious diseases are in extreme environments and lack access to quality hospital-level clinical care. Further, traditional diagnostic tools require experts to read and interpret results, limiting their applications at the point of care in arid conditions. AI and machine learning can bridge the gap in such regions, enabling translation of diagnostics to individual hands for rapid response, precision health, and personalized medicine, but three main challenges hinder its effective application: limited clinical training data, computational intensity, and lack of on-device AI systems. This project aims to develop novel, optimized machine learning algorithms that can be implemented on a device without reliance on the cloud, particularly for implementation in a rapid infectious disease diagnostic tool developed in the Tadesse Lab at MIT. This approach has the potential to significantly impact prevention and preparation against the next pandemic.
- “Learning Latent Human Needs for Real-World Generative Design of Physical Systems”
Caitlin Mueller, Departments of Architecture and Civil and Environmental Engineering
Maria Yang, Department of Mechanical Engineering
Generative deep learning models can return convincing design artifacts (images, prose, 3D models) from semantic prompts, presenting a new paradigm for the process of designing. Instead of explicitly specifying a concept (or a system for creating a concept), designers can provide much more abstract instructions to computational systems and obtain potentially valuable results. However, while language-based generative models are exciting in their novelty, they are still far from expressing and responding to the actual needs, desires, and knowledge of human users and designers of physical products, systems, objects, and spaces. This project proposes to advance the use of AI to more fully identify, express, and integrate latent user needs and tacit designer knowledge into creative design processes based on generative modeling.
- “Studying How Blind/Low-Vision People Can Validate and When They Prefer Automated Chart Captions”
Arvind Satyanarayan, Department of Electrical Engineering and Computer Science and Computer Science and Artificial Intelligence Laboratory
Automated chart captions for blind/low-vision people present a conducive petri dish for studying broader open questions in generative AI and human-AI interaction as it offers a variety of structured and formal domain-specific representations to bring to bear on each question. In particular, by leveraging formal specifications of the chart’s construction as well as the input dataset, this project will build novel interactive scaffolding to allow blind/low-vision people to explicitly and implicitly verify the insights conveyed in an automatically generated caption. Similarly, low-and high-level taxonomies of analytic tasks readers perform with charts and captions will provide an experimental design framework to empirically test when blind/low-vision people might prefer a caption over digging into the data for themselves — and how they might wish to integrate both modalities.
- “Advanced Optimization for Neural Network Training with Differential Privacy”
Rahul Mazumder, MIT Sloan School of Management and Operations Research Center
Protecting privacy of training data for machine learning models has become even more important with the invention and rising proliferation of current high-capacity generative large language models (LLMs). Differential Privacy (DP) is a well-established framework for reasoning about data anonymization. Modifying the training process (DP-training) of models like LLMs to provide differential privacy guarantees is a theoretically justified way of protecting training data. At the same time, applying DP-Training remains hard in practice due to privacy-utility-computation tradeoffs. While part of this tradeoff is inherent to data protection, some of the utility loss is due to imperfect methods of DP-Training and optimization of the DP-Trained models. This project will explore several directions that can improve DP-Training utility-privacy-computation tradeoffs.
- “Using Google’s Newly Discovered Crystals to Decarbonize the Economy”
Ju Li and Bilge Yildiz, Department of Nuclear Science and Engineering
Jesús del Alamo, Department of Electrical Engineering and Computer Science and Microsystems Technology Laboratories
Based on historical trends, physical technologies such as chemical and materials production, energy generation, transmission, transportation, and construction tend to evolve significantly more slowly than information technology. Shifting dependence from fossil fuels to cleaner alternatives will require a herculean effort, but it is a crucial task that must be undertaken to secure a sustainable future. This project seeks to leverage Google’ accomplishments in the discovery of novel stable crystals to allow rapid insertions of new materials technologies into the economy. Utilizing the large database of new thermodynamically stable crystals found by Google Brain into different industrial applications, the project team will screen these new crystals’ figure-of-merit for batteries, electrolyzers, superconductors, hydrogen-storage materials, and energy-efficient computing devices, and then rapidly synthesize and characterize some crystals experimentally at MIT.