The following is a guest blog post by CCC AI and Robotics Task Force Co-Chair Greg Hager from Johns Hopkins University.
Artificial intelligence (AI) has emerged into the public view as an important frontier of technological innovation with potential influences in many realms. Many recent symposia and workshops including AI for Social Good, Computing Research: Addressing National Priorities and Societal Needs, and Discovery and Innovation in Smart and Pervasive Health have highlighted both the progress and opportunities for AI and its potential to contribute to new products, services, and experiences.
However, we should not lose sight of the fact that fielding real-world systems that realize these innovations will also drive significant advances in virtually all areas of computing, including areas that are not traditionally recognized as being important to AI research and development.
To highlight these synergies, the AI and Robotics Task Force, led by Greg Hager from Johns Hopkins and Eric Horvitz from Microsoft Research and their current members Randy Bryant from Carnegie Mellon University, Maja Matarić from the University of Southern California, and Vasant Honavar from Pennsylvania State University, has just released a white paper for the community called Advances in Artificial Intelligence Require Progress Across all of Computer Science.
This report suggests several promising areas of interaction between AI and the broader computer science research community. These areas include:
Computing systems and hardware. There are opportunities ahead for leveraging innovations in computing systems and hardware. Directions include the development of methods for speeding up core computational procedures employed in AI systems, such as the methods used to train and to execute classification for perceptual tasks using convolutional neural networks. Opportunities include new approaches to parallelism, smart caching, and uses of specialized hardware like FPGAs to lower costs of computation and to meet the demands and robustness needed with AI applications.
Theoretical computer science. AI was built on theoretical work based on the mathematics of computability in the early 20th century by Turing, Church, and others. AI challenges have long posed and faced combinatorial challenges and has made use of results on the performance and precision of approximation procedures. There are continuing opportunities for the formal study of hard challenges in AI with tools and techniques developed in the realms of analysis of algorithms, including efforts in combinatorics, computational complexity theory, and studies of computability.
Cybersecurity. AI systems are being developed for high-stakes systems in such areas as healthcare and transportation. These systems are also bringing to the fore new attack surfaces that need to be understood and protected. Directions include understanding and hardening systems to whole new categories of attack including, “machine learning attacks,” where clever adversarial procedures are employed to inject data into systems that will confuse or bias them in their intended operation. AI systems frame new challenges that will require advances in security that address the new attack surfaces to ensure that they are safe, reliable, robust and secure against malicious attacks.
Formal methods. Formal methods can play a critical role in defining and constraining AI systems, so as to ensure that their behavior conforms to specifications. Efforts include methods for doing formal verification of programs and also to perform real-time verification of systems through new kinds of monitoring. Formal methods are promising approaches to ensuring that programs do not take actions beyond specified goals and constraints.
Programming languages, tools, and environments. New programming languages, tools, and programming environments can help engineers to build, test, and refine AI systems. Higher-level languages can offer engineers and scientists new kinds of abstractions and power to weave together multiple competencies, such as a vision, speech recognition, and natural language understanding so as to be able to develop and debug programs that rely on the close coordination of multiple AI analytical pipelines.
Human-computer interaction. The key challenges with AI frame numerous opportunities in the broad realm of research in human-computer interaction (HCI), an important area of computer science research. Efforts include methods for explaining the results of AI systems to people, allowing people to work interactively with AI systems (e.g., interactive machine learning), that help with the specification, encoding, and understanding of the implications of different policies, values, and preferences assumed by automated systems, and supporting new kinds of human-AI collaboration, including mixed-initiative interaction and augmenting human cognition.
There is a growing and compelling imperative to leverage the advances in AI and automation to improve human lives in many ways. The path toward a balanced portfolio of capable, safe, and transparent AI-based systems will draw on a broad spectrum of computing ideas and principles, and is likely to become a driver for new advances in computing. By embracing the promise of AI, we believe that many areas of computer science will not only be advanced, but will also allow AI to address important opportunities and do so in a way that is safe, reliable, and effective.
The AAAI 2017 Spring Symposium on AI for Social Good (AISOC) (co-sponsored by Association for the Advancement of Artificial Intelligence (AAAI) and the Computing Community Consortium (CCC)) will be on March 27-29, 2017 at Stanford University. The symposium will focus on the promise of AI across multiple sectors of society. See more information here.
.
