Amazon and MIT have named the recipients of the 2023 gift project and fellowship awards granted as part of the Science Hub.
The Science Hub — administered at MIT by the Schwarzman College of Computing — is a collaboration between Amazon and MIT that supports leading-edge research, education, and outreach in developing technologies such as machine learning and robotics.
“This year’s Science Hub projects are addressing some of the most challenging problems in machine learning and robotics. Our researchers are taking a hard look at relevant issues that have emerged, including those within autonomous systems, large language models, and data, and are committed to uncovering practical solutions,” said Sertac Karaman, MIT head of the Science Hub, director of the Laboratory for Information and Decision Systems (LIDS), and professor of aeronautics and astronautics at MIT. “I’m also excited to welcome our new fellows, who will conduct their own independent research projects in AI and robotics.”
“We are excited to support research in next-generation large-language-model technologies,” said Shiv Vitaladevuni, director of applied science in Alexa. “These gifts will support development of more reliable LLMs and leveraging LLMs for data curation. Leveraging LLMs for data curation will increase developer productivity to summarize the contents of databases and address issues such as noisy, missing or biased data. Supporting the science community is our way to give back to academia for long-term benefit to society.”
The research projects further the goals of the hub, which are to ensure that the benefits of new technology are broadly shared through education and research, as well as to expand participation in research from a wide variety of scholars and other innovators.
“We are committed to supporting the best robotics and machine learning science,” said Jeremy Wyatt, director of applied science at Amazon Robotics. “These gifts will support breakthrough work in the application of machine learning to robotics. The projects will advance the safe deployment of machine-learned models in autonomous systems and the realization of simultaneous localization and mapping on energy-efficient devices. Giving back to academic science is just one of the ways that we seek to live up to our principle that success and scale bring responsibility.”
A committee of members from both Amazon and MIT selected the four research projects, which are detailed below.
“Online safety monitoring for AI-enabled robot autonomy” — Navid Azizan, Esther & Harold E. Edgerton Career Development Assistant Professor
“Deep neural networks in autonomous systems, such as robots, can be untrustworthy on inputs that are significantly different from their training dataset—for instance, when a robot encounters unexpected objects blocking its path or interacts with a human who deviates from the robot’s expected actions. Such scenarios can lead to accidents or halts in operations. To ensure safe deployment, autonomous systems should detect anomalies or out-of-distribution data points and react to them by delegating the decision or switching to a safe baseline policy. This could prevent blockages and accidents while improving system efficiency. We propose to investigate how to endow deep neural networks in autonomy stacks with the capability to detect anomalies efficiently and effectively, which can in turn continually improve the model by learning from anomalies over time.”
“Controlling large language models with symbolic structures” — Yoon Kim, assistant professor of electrical engineering and computer science
“Despite their impressive capabilities, large language models remain difficult to control. This proposal seeks to develop mechanisms for controlling LLMs through symbolic grammars. Given a pretrained model, our approach defines a probabilistic grammar whose nonterminal symbols are explicitly related to the pretrained model’s distribution over partial outputs. These symbols provide an interface with which to interact with (and place constraints on) the language model, which can achieve more explicit constrained generation than prompt-based approaches (which cannot guarantee that LLM outputs respect constraints). We propose to apply this approach on zero- and few-shot generation tasks where respecting output constraints on the target side is crucial for deployment (e.g., semantic parsing, translation of medical notes).”
“DataCore: A foundational model for enterprise data curation” — Samuel Madden, MIT College of Computing Distinguished Professor of Computing
“Modern data-intensive applications are characterized by a need to combine and query a variety of datasets, ranging from internal documentation, to partially structured data such as logs, to tabular data in databases, to machine learning modes and outputs. In many organizations, such data is spread across the enterprise and often inconsistently structured, incomplete, and unlinked. To conduct data analytics on this poorly maintained data, data scientists must go through a data curation process to find, merge, and clean datasets. Despite years of research on this problem, many data scientists still report spending 80% or more of their time on such problems. We aim to address these shortcomings by building a data-curation-native foundation model that has all the merits of foundation models in NLP as well as other key features for handling big, structured tabular datasets common in enterprise data. Our goal is to develop a generic model that can effectively serve various data curation tasks over structured data with state-of-the-art or better performance, which, like ChatGPT, does not require a large amount of domain-specific training but can be guided through interactive prompting.”
“Enabling memory-efficient SLAM for energy-constrained devices” — Vivienne Sze, associate professor of electrical engineering and computer science
“We propose to co-design algorithms and hardware for simultaneous localization and mapping (SLAM) that are efficient, robust, and accurate all at the same time. Energy-constrained devices like smart phones, AR/VR headsets, and pill-size medical robots are set to make significant contributions to a diverse set of applications. However, these devices have limited battery capacity, which restricts the available energy for sensing and computation. Still, to operate safely, devices should perform fundamental tasks such as determining their location without GPS (localization) and creating a representation of obstacles in their environment (mapping). Existing algorithms for these tasks require too much memory and energy overhead and also struggle in the presence of sensor noise and insufficient sensor modalities. Thus, implementing autonomy on energy-constrained devices requires both the design of robust and efficient localization-and-mapping algorithms and specialized energy-efficient computing hardware.”
2023 Fellowships
The following doctoral students will receive funding to pursue independent research projects in robotics and AI. Students will have an opportunity to participate in paid summer internships at Amazon where they can work directly with Amazon researchers to gain valuable industry insight and experience.
Sirui Li, PhD candidate, social and engineering systems and statistics
Li received her bachelor’s in computer science and mathematics from Washington University in 2019 and was one of the five Class of 2019 valedictorians from the Washington University School of Engineering. Li is especially interested in Bayesian methods, social networks, and applications of machine learning to economics and political science. Li works with Cathy Wu, the Gilbert W. Winslow Career Development Assistant Professor, Civil and Environmental Engineering, on solving the ride-sharing problem using graph neural networks and reinforcement learning. Li’s long-term research goal is to improve algorithms/heuristics that solve hard social problems in city planning and mechanism design.
Yue Meng, PhD candidate, aeronautics and astronautics
Meng is a fourth-year PhD student at MIT AeroAstro, working in the Reliable Autonomous Systems Lab. Meng’s research topic is using machine learning techniques for safe and robust robot control. Before that, Meng was an AI resident at the IBM Thomas J. Watson Research Center. He earned a master’s in electrical and computer engineering at the University of California, San Diego, and received a bachelor of science degree from Tsinghua University in the Department of Automation.
