Research

Artificial intelligence involves providing capabilities in computers or systems embedded with computing the ability to perform tasks that have traditionally required human involvement and intelligence. In recent years, for a narrow set of tasks such as object detection in images, machines have exceeded human capability. However, significant gaps exist to extend this success of AI to a broad set of tasks and variety of science and engineering applications. CAFE will lead research at the convergence of foundational and applied AI across analytic methods and application domains to exploit breakthroughs at one frontier that can apply at others.

CAFE is organized into two research thrusts.

Foundational AI

Research in foundational AI seeks to develop theory and methods that are common across application domains. It will encompass the design of new algorithms, knowledge representation and abstractions, optimization strategies, learning approaches, human-machine interaction, and machine-machine interactions. The thrust will specifically leverage cross-disciplinary expertise to address issues such as understanding the foundational aspects of machine learning by working at the interface of machine learning and mathematical statistics. The outcome of this research will result in advances such as unsupervised learning without the need for large, annotated data sets and the ability to learn multiple tasks and objectives at the same time, provide reasoning for decisions, understand causal relationships, and provide a balance between preserving privacy and utility of shared data. CAFE will bridge the challenges associated with the increasing reliance on intelligent systems for the secure collection, handling, and communication of personal data and associated actions initiated by machines in human-machine interaction contexts. This thrust will also encompass efficient hardware and software systems that realize efficient AI systems. Advances in AI will continue to be dependent on hardware optimized for AI algorithm and related efforts toward reducing the amount of energy expended. Recently, training of a natural language processing model was estimated to emit more than 626,000 pounds of carbon dioxide equivalent — nearly five times the lifetime emissions of the average American car. Further, the need for embedding intelligence in the power-limited internet of things is increasing the need for energy efficiency. CAFE will explore brain-inspired and new technology driven architectures such as neuromorphic processors and domain-specific accelerators for addressing these needs. This research will aim to achieve 100-1000X enhancement in power efficiencies compared to existing systems. As hardware systems advance and can support learning and inference at the edge, theoretical innovations such as distributed learning and cooperative inference will become essential.

Applying AI to Engineering Systems

This thrust focuses on use-case specific augmentation or tuning of the foundational AI elements. Successful application of AI will depend on the ability to integrate domain-knowledge and needs into the AI system. CAFE will focus on applying AI to engineered systems and in enabling scientific discoveries. The center will leverage strengths of its faculty in applying machine-learning techniques towards materials discovery, earthquake prediction, smart manufacturing, additive manufacturing, intelligent transport systems, soft robotics and brain-machine interfaces. One of the goals of this thrust is to help rapid adoption of AI technologies to address a variety of challenges by creating a shared repository of methodologies and tools and a peer group of users. Additionally, the variety of application domains will drive innovations in the foundational AI thrust such as transfer learning from one domain to another. This thrust will also help address challenges of creating robust data sets for training by bringing together a critical mass of researchers and develop AI approaches that are robust to experimental and measurement variations of acquired training data. One key challenge for adoption of AI techniques toward applied problems is the rapid evolution in foundational AI advances. The algorithms and techniques employed are rapidly evolving, making the risks of obsolescence high. The interactions facilitated by CAFE among researchers will help reduce the time to deploy new AI frameworks and methodologies and increase the interdisciplinary interactions that will mitigate the risk of lost opportunity and missed generalizations across disciplines and domains. This can help more rapid explorations in various application domains using the power of AI.