MIT researchers have pioneered a groundbreaking approach to artificial intelligence that dramatically simplifies entry for newcomers while simultaneously empowering experts to push the boundaries of the field.
In a revolutionary research paper unveiled at the prestigious Programming Language Design and Implementation conference, the research team introduces "Gen," an innovative probabilistic-programming system. This cutting-edge platform enables users to construct models and algorithms across diverse AI application domains—including computer vision, robotics, and statistics—without wrestling with complex equations or laboriously crafting high-performance code. Gen additionally empowers seasoned researchers to develop sophisticated models and inference algorithms for prediction tasks that were previously impractical to implement.
The research paper showcases how a concise Gen program can effectively determine 3-D body positions—a challenging computer-vision inference task with critical applications in autonomous systems, human-machine interactions, and augmented reality. This program seamlessly integrates components for graphics rendering, deep-learning, and various probability simulation techniques. This fusion of diverse methodologies delivers superior accuracy and processing speed compared to previous systems developed by the team.
Thanks to its user-friendly design and, in certain applications, automated functionality, Gen accommodates users across the expertise spectrum. "A primary objective of this work involves making automated AI more accessible to individuals without extensive computer science or mathematical backgrounds," explains lead author Marco Cusumano-Towner, a doctoral student in the Department of Electrical Engineering and Computer Science. "We're equally focused on enhancing productivity, enabling experts to rapidly prototype and iterate their AI systems."
The researchers further demonstrated Gen's capacity to streamline data analytics through another program that automatically generates sophisticated statistical models typically employed by experts for analyzing, interpreting, and predicting underlying patterns in data. This advancement extends the team's previous research that allowed users to uncover insights into financial trends, air travel patterns, voting behaviors, and disease propagation with minimal code. Unlike earlier systems requiring extensive manual coding for accurate predictions, Gen automates much of this process.
"Gen represents the first system that delivers the necessary flexibility, automation, and efficiency to excel across these vastly different domains in computer vision and data science while maintaining state-of-the-art performance," notes Vikash K. Mansinghka '05, MEng '09, PhD '09, a researcher in the Department of Brain and Cognitive Sciences who leads the Probabilistic Computing Project.
Collaborating with Cusumano-Towner and Mansinghka on the paper are Feras Saad '15, SM '16, and Alexander K. Lew, both CSAIL graduate students and members of the Probabilistic Computing Project.
Optimizing All Approaches
When Google released TensorFlow in 2015, this open-source library of application programming interfaces (APIs) revolutionized the field by enabling both beginners and experts to automatically generate machine-learning systems with minimal mathematical expertise. While now widely adopted and helping democratize certain aspects of AI, TensorFlow primarily focuses on deep-learning models that are both resource-intensive and limited compared to the broader potential of artificial intelligence.
Numerous alternative AI techniques exist today, including statistical and probabilistic models, along with simulation engines. Some existing probabilistic programming systems offer sufficient flexibility to encompass multiple AI approaches, but they suffer from significant performance limitations.
The MIT team endeavored to create a system that harmonizes the most desirable attributes—automation, flexibility, and speed—into a single cohesive platform. "By achieving this integration, we can help democratize this much broader collection of modeling and inference algorithms, similar to how TensorFlow transformed deep learning," Mansinghka explains.
Within probabilistic AI, inference algorithms process data and continuously adjust probabilities based on new information to generate predictions. This iterative process ultimately produces a model capable of making predictions on previously unseen data.
Building upon concepts from their earlier probabilistic-programming system, Church, the researchers integrated several custom modeling languages into Julia, a general-purpose programming language also developed at MIT. Each modeling language receives optimization for specific AI modeling approaches, enhancing versatility. Gen further provides high-level infrastructure for inference tasks, employing diverse methods including optimization, variational inference, specific probabilistic techniques, and deep learning. Additionally, the researchers implemented performance-enhancing modifications to ensure efficient execution.
Real-World Applications
External organizations have already begun leveraging Gen for their AI research initiatives. Intel, for instance, is collaborating with MIT to employ Gen for 3-D pose estimation using depth-sensing cameras in robotics and augmented-reality systems. Similarly, MIT Lincoln Laboratory is partnering on applications for Gen in aerial robotics for humanitarian assistance and disaster response scenarios.
Gen is being implemented in ambitious AI projects within the MIT Quest for Intelligence. Notably, Gen plays a central role in an MIT-IBM Watson AI Lab project, as well as the U.S. Department of Defense's Defense Advanced Research Projects Agency's ongoing Machine Common Sense project, which aims to model human common sense at the level of an 18-month-old child. Mansinghka serves as a principal investigator on this project.
"With Gen, researchers can for the first time easily integrate multiple AI techniques. It will be fascinating to discover what becomes possible now," Mansinghka observes.
Zoubin Ghahramani, chief scientist and vice president of AI at Uber and a professor at Cambridge University, who wasn't involved in the research, states, "Probabilistic programming represents one of the most promising frontiers in AI since the emergence of deep learning. Gen constitutes a significant advancement in this field and will contribute to scalable and practical implementations of AI systems based on probabilistic reasoning."
Peter Norvig, director of research at Google, also commended the work. "Gen enables problem-solvers to utilize probabilistic programming, thereby adopting a more principled approach to problems without being constrained by the decisions made by the system designers," he notes. "General-purpose programming languages have succeeded because they make tasks easier for programmers while simultaneously enabling them to create entirely new solutions to efficiently address novel challenges. Gen accomplishes the same for probabilistic programming."
Gen's source code is publicly accessible and will be featured at upcoming open-source developer conferences, including Strange Loop and JuliaCon. The research receives partial support from DARPA.
