Tutorial

Ideation, or the ability to generate new ideas from the mental substrate of semantic knowledge, is a key attribute of human intelligence. However, the mechanisms underlying this process are still open to debate. Most experimental work on semantic cognition has focused on specific response tasks such as recall and recognition rather than on productive tasks like ideation. However, a significant body of information on the latter has accumulated through the study of brainstorming in individuals and groups, as well as indirectly through work on the effects of priming in response tasks. Recently, we have developed a model that tries to synthesize these behavioral results with current understanding of the neural level processes underlying working memory, cognitive control, reward mechanisms and learning. The goal is not only to understand the neural basis of ideation, but also to potentially obtain novel computational models for thought and creativity. Such models could be useful in applications requiring ceativity, including design, artistic composition, and brainstorming.

The model is based on the fundamental assumption that ideas emerge from the natural dynamical interaction between patterns of neural activity in brain regions encoding semantic information at different levels. Ideas are modeled as combinations of co-active concepts in a recurrent network whose itinerant dynamics represents the a search process in idea space. This search is shaped by a dynamic bias from a working memory/cognitive control system, which is modeled as a modular recurrent network with context-specific hierarchical attractors. The ideas generated during the search are evaluated by a critic with domain knowledge, and this evaluation is fed back to modulate and reconfigure the attractors shaping the search. During training, the system embeds semantic knowledge in its synaptic weights, so that the dynamics at both the idea and attentional levels takes this knowledge.

1. Recall: Producing previously learned ideas automatically in familiar contexts.
2. Generalization over Ideas: Generating appropriate new ideas in familiar contexts.
3. Generalization over Contexts: Generating appropriate ideas in unfamiliar contexts.

Human ideation often occurs in social situations -- notably in group brainstorming sessions that are a staple of organizational practice. Behavioral studies have shown that many social factors influence the ideational productivity of groups. We have implemented a multi-agent model of group brainstorming, where each agent is defined by a simplified (but still neural) version of the individual ideation model described above. The agents interact in various spatial and temporal configurations, and can differ in both their underlying knowledge and their style of ideation. The model is used to study the effect of interaction protocols and agent diversity on the productivity of brainstorming.

Overall, the approach taken sees cognition and behavior as emergent properties of the brain-body system, and explores the idea that the neural structures and mechanisms underlying thought and action are essentially similar, albeit differing in some of the brain regions involved.

1. An overview of semantic cognition, ideation and brainstorming data.
2. Detailed description and demonstration of the dynamical idea generation model.
3. Discussion of the attractor dynamics and learning strategies proposed to underlie the ideation process.
4. Exploring broader links between the dynamics of cognition and behavior.
5. A multi-agent model of brainstorming in groups of interacting individuals, each modeled by a simplified version of the dynamical model.

Ali A. Minai received his Ph.D. in Electrical Engineering from the University of Virginia in 1991, with a dissertation on feed-forward neural networks. From 1991-93, he was a post-doctoral research associate in the laboratory of Prof. William B at the University of Virginia, working on computational models of sequence learning in the hippocampus. In 1993, he joined the Department of Electrical & Computer Engineering at the University of Cincinnati, where is is currently an Associate Professor and Director of the Complex Adaptive Systems Laboratory. He is also a faculty member of the Neuroscience Graduate Program at the University of Cincinnati. Dr. Minai's research interests include neural networks, computational neuroscience, computational models of cognition and action, biomorphic systems, distributed multi-agent systems, and self-organized networks. His research has been funded by NSF and the Air Force Research Lab. He has published more than 80 refereed papers.

Dr. Minai is co-editor of five books in the area of complex systems, including "Complex Engineered Systems: Science Meets Technology", published by Springer Verlag in June 2006. He has been Secretary of the International Neural Networks Society (INNS) since 2006, and served as an Associate Editor of the IEEE Transactions on Neural Networks from 2001 to 2008. He was also a member of the IEEE-CIS Neural Networks Technical Committee in 2004-2005. He has served on the Executive Committee for the International Conference on Complex Systems since 2000, helping to organize 5 high-profile conferences. He was Workshops Chair for IJCNN 2007 and Tutorials Chair for IJCNN 2009, and has served on the program committees of several other international conferences, including ICANN 2009. He has also served on several National Science Foundation review panels.

Dr. Minai is a senior member of the IEEE, and the International Neural Networks Society, and a member of the Society for Neuroscience, Sigma Xi, Eta Kappa Nu, and Tau Beta Pi. He served as Treasurer of Sigma Xi's University of Cincinnati Chapter from 2000 to 2004. He received the William E. Restemeyer Teaching Excellence Award in 1995 and the Eta Kappa Nu Outstanding Professor of the Year Award in 2002 from the ECECS Department at the University of Cincinnati. In 2004, he was recognized by the University of Cincinnati's "Future of Learning: Addressing Issues of Diversity" initiative.