This project is to host a workshop focused on identifying frontiers for collaborative research integrating mathematical and computational modeling of human cognition with machine learning and machine intelligence. The workshop will bring together leading researchers who are working at the intersection of the fields. The goal is to initiate dialogue between people working within different computational frameworks, to identify points of synthesis where new scientific insights or breakthroughs are most likely, and to identify gaps and obstacles to collaboration between the two communities. A summary report will be produced that details the products of the discussion, highlighting prospects identified for future collaborative research.
At a workshop sponsored by the National Science Foundation, 18 distinguished researchers from the fields of Cognitive Science (CS) and Machine Learning (ML) met in Arlington, VA in May 2013 to discuss computational approaches to cognition in human and artificial systems. The purpose of the workshop was to identify frontiers for collaborative research integrating (a) mathematical and computational modeling of human cognition with (b) machine learning and machine intelligence. The researchers discussed opportunities and challenges for how the two fields can advance each other and what sort of joint efforts are likely to be most fruitful. There are several reasons to believe that theories of human cognition and of machine intelligence are currently in position to greatly benefit from each other. The mathematical and computational tools developed for designing artificial systems are beginning to make an impact on theoretical and empirical work in CS, and conversely CS offers a range of complex problems that challenge and test ML theories. ML systems would likely be more successful if they were more flexible and adaptive, like human cognition, and if they were built on richer representations that (in some sense) embody meaning or understanding as opposed to black-box statistics. The synthesis is also timely because CS researchers are starting to work on larger datasets, and a major focus of ML in the last decade has been to develop tools that can succeed in complex, naturalistic situations. The attendees concluded that the fields of cognitive science and machine learning, having been split for many years, are moving into a period of greater interaction and synergy. Future collaboration between CS and ML could open up new theoretical territory and produce major breakthroughs relevant to technology and society at large. CS and ML are united by the shared goal of developing a computational understanding of human-level perception and cognition. However, they have tended to focus on different aspects of this problem—ML emphasizing powerful statistical algorithms that scale to complex stimuli or tasks, and CS emphasizing structured and flexible representations that can apply to multiple tasks. The strengths of the two fields are thus naturally complementary. The primary challenges to collaboration may lie in the pragmatics of how the two fields work. What counts as a useful advance in CS is not necessarily considered so in ML. For cognitive ideas to have a greater impact in ML might require demonstrating that they can provide quantitative improvements on objective benchmarks or that they suggest new benchmarks that can push theory forward. The two fields’ infrastructures for training and dissemination are also largely separate. Future funding for conferences, workshops, graduate courses, and summer schools might go a long way toward creating a new generation of interdisciplinary researchers. The workshop generated several ideas for specific collaborative research areas where the next breakthroughs might occur, including: methods for better, deeper generalization; flexibility and self-programming; planning and acting in complex dynamic environments; transitioning from learning algorithms to complete learning models; exploring the theoretical limits of learning; using big data to understand cognition and behavior; and applying CS and ML together to explain the brain. The discussion also identified potential benefits to society from research integrating CS and ML, including a new generation of intelligent artificial systems, improved decision-making both for individuals and in public policy, advances in adaptive and personalized education and training, and new computational frameworks for understanding the brain. The conclusions from the workshop are presented in detail in the Workshop Report, available from NSF or at http://matt.colorado.edu/compcogworkshop/report.pdf.
