Understanding Brainstorming and Creativity Through Behavioral Experiments and Neural Models
Associate Professor, Department of Computer Science
Associate Professor, Department of Psychology
Brainstorming, as first defined by Osborn (1957), is a set of guidelines for groups trying to come up with novel and creative ideas on a particular problem. Group brainstorming sessions are popular, the assumption being that people are more likely to come up with ideas in a group that they would not have come up with on their own. Many people also report that for many tasks, such as brainstorming, they prefer to work in a group rather than work alone. So, it is quite surprising to learn that under controlled experimental conditions, groups of people trying to generate creative solutions to a problem generally come up with far fewer ideas than an equal number of people working alone on the same problem (approximately half the number of ideas, in fact) (Diehl and Stroebe, 1987; Paulus and Dzindolet, 1993). There are a number of reasons for this. People may be afraid to express seemingly creative ideas for fear of being criticized, or they may get lazy and “free ride,” assuming other group members will pick up the slack. Another reason, called “blocking,” is due to the group setting where people compete for speaking time, may be overshadowed by others, or forget their ideas or decide not to express them while waiting their turn (Diehl and Stroebe, 1987; Paulus and Dzindolet, 1993). On the other hand, other reasons, such as increasing motivation and accountability of performance (Larey and Paulus, 1999), enhance brainstorming in both individuals and groups. It is believed that interaction with others’ ideas helps people unveil remote associations that could have not been accessed alone. For example, in electronic group brainstorming, participants contribute ideas via a computer network and see ideas from others appearing on their monitors. This setup has been shown to reduce the blocking effect observed in classic group brainstorming (Paulus, 2006; DeRosa, 2007). To understand the type of others’ ideas that lead to more original ideas, many experiments substitute others’ ideas with controlled, externally provided hints. For example, exposing brainstormers to hints during individual or group brainstorming was shown to increase the number of ideas, with better results if hints were spread out during the session as opposed to being presented all at the beginning (Coskun et al., 2000).
Nonetheless, group brainstorming sessions are popular and are utilized quite often. One advantage, as mentioned above, is that group members can share their expertise, different people know different things, and, in a group, that knowledge can be pooled. Another advantage of groups could be that people can be helped by being exposed to other people’s ideas: something someone else says could lead you to think of something you wouldn’t have thought of by yourself, or you might generate a novel idea by combining one of your ideas with someone else’s, or you might use someone else’s idea as an analogy or metaphor to stimulate an idea.
Although it is far from clear that groups and teams are always effective, there are many situations in which collaboration is unavoidable. In situations that require a variety of expertise, for example, working as a team is a necessity. As scientists study increasingly complex problems, collaboration among scientists is increasingly common.
Interestingly, studying teamwork and collaboration is one scientific problem that often requires a team of scientists. Hofstra Professors Simona Doboli and Vincent Brown are part of a team that has received funding from the National Science Foundation to perform a detailed study of the ways in which working in a team has both negative and positive effects on creative idea generation, or brainstorming. Other members of the team are Professors Paul Paulus and Daniel Levine at the University of Texas at Arlington and Professor Ali Minai at the University of Cincinnati. The purpose of this collaborative research is to advance the understanding – through computational modeling and behavioral experiments – of conditions and neural mechanisms that enhance the production of novel ideas. This is the first such interdisciplinary effort to study brainstorming.
Professors Brown and Doboli arrived at brainstorming research by making analogies with their past research interests. Dr. Brown’s research area was in mechanism of attention in the brain and, in particular, in a type of neural processing called lateral inhibition. Lateral inhibition refers to the situation where an active neuron or group of neurons suppresses or inhibits other neurons or groups of neurons. This is useful for paying attention to important events or objects around you (for example, someone talking to you) while suppressing or inhibiting potential distractions (such as the radio). So tending to one thing blocks your attention to something else. Listening to Dr. Paulus’ talk on brainstorming, while a faculty member at University of Texas at Arlington, Dr. Brown made the analogy between the ways people interfere with each other’s thought processes when speaking and interacting in a group, a problem that previous researchers labeled as “blocking,” and lateral inhibition and attention processes he was studying in the brain. In other words, if one person in a group is expressing ideas, that may suppress the ability of other group members to express their ideas. One active person suppresses or inhibits other people in the group, analogous to the way in which lateral inhibition is the process by which one active group of neurons suppresses the activity of another group of neurons. The usefulness of this analogy led Professors Paulus and Brown to begin collaborating on studying the disadvantages – as well as the potential advantages – of brainstorming in a group.
One important result of their collaboration is the cognitive associative model of brainstorming (Brown et al., 1998). The model represents semantic knowledge as a network of categories, and ideas as a sampling process of the pool of available categories, where a category is seen as a pool of semantically related concepts or ideas. For example, a typical problem used in brainstorming experiments is that of asking participants to generate ideas to improve university life. Ideas can be grouped in similar semantic categories, such as ideas related to parking, faculty, cafeteria, and so on. The model has been very successful in explaining individual and group brainstorming experiments (Coskun et al., 2000) and in predicting factors that would enhance brainstorming productivity. The model is able to emulate attention to others’ ideas, different types of brainstormers (e.g., divergent, those who jump easily from one category to another; and convergent, those who switch categories less often), different types of categories (e.g., high-accessible, those that are more frequently used to generate ideas from, e.g., parking; and low-accessible, those that are rarely sampled for ideas, e.g., landscaping). The main shortcoming of the associative model is the abstract representation of individual ideas, and, hence, its inability to study the quality and novelty of ideas generated. New ideas are thought to be novel conceptual combinations of old concepts (Mednick, 1962; Boden, 1995). The cognitive associative memory model (Brown et al., 1998) cannot explicitly generate novel ideas and thus cannot be used to study the factors that influence the production of such ideas.
Before Professors Doboli and Brown met at Hofstra, Professor Doboli’s research was on developing neural models for representing context in the brain, for example, how same words, or same visual scenes, in different contexts, activate distinct populations of neurons in the brain – hence they are represented differently. For example, the meaning of “chips” in the cafeteria is totally different than the meaning of “chips” in a computer science class. In her doctoral work, she studied how rats represent in the brain different environments as a complex function of not only what they could see, touch or smell, but of the task they had to perform, or of the entry point in the environment. A new concept in her modeling work was that of latent attractors (Doboli, Minai and Best, 2000), where one such attractor is a pool of neurons whose activity continuously encodes the changing position of the animal in the environment. Listening to Professor Brown’s talk about the cognitiveassociative model and brainstorming, Professor Doboli made the analogy with the latent attractor model. Each latent attractor, or context, can represent a distinct semantic category, and ideas can be encoded as different activity patterns within the pool of neurons in a latent attractor, in the same way as the position of the animal was encoded. Professors Doboli and Brown started to work together on a neural model of idea generation that could represent distinct ideas and could account for novel ideas as combinations of old ones.
Neural Models of Idea Generation
The work done so far by our group as part of the National Science Foundation grant led to a new, full-scale neural model of idea generation (Doboli and Minai, 2007; Iyer et al., 2007; Iyer et al., 2008). The model encodes semantic knowledge at different levels of abstraction, including categories of semantically related concepts, individual concepts such as words or combination of words, and features of concepts. Categories, concepts and features are all represented as neural units. Connections between different units in different levels or layers, as well as within the same layer, determine which combinations of concepts are activated depending on the context of the brainstorming problem. Ideas are expressed as combinations of active concept units. They are generated by a search process guided by the problem description or context, and are constrained to contain a minimum set of desirable response features. An internal evaluation process, or critic, evaluates each idea generated by comparing the set of active features with the set of desirable features within that context. If a minimum criteria is not passed, then the search space (i.e., the concepts in the active categories) is widened to include other categories. The order in which categories are added to the search space depends on the relationships between them and the categories already in the search space. If the minimum criteria for idea quality are met, then the search space is narrowed slowly to find the minimum subset that produces the best ideas. More importantly, the whole system with the dynamic, adaptive search process is implemented with mechanisms similar to the ones used in the brain. It is the first such model to study idea generation. The model is able to find good ideas in different contexts, both familiar and unfamiliar.
A more detailed neural model focusing more closely on the concept layer from the full-scale model described above was developed by Professors Doboli and Brown (Doboli, Brown and Minai, forthcoming, 2009). The goal of this model is to better understand the internal dynamics of the idea generation process and the factors that influence the production of new ideas. The conceptual neural model represents only concepts and the relationships between them. Mathematically, the model was inspired by the modular of the cortex and models of interacting populations of excitatory and inhibitory neurons. The dynamics of idea generation in the conceptual neural model are determined by: (a) the connectivity pattern and strength of existing semantic relations between concepts, (b) the sequence of previously generated ideas, and (c) the modulation of activity controlling the semantic distance between co-active concepts.
Semantically related concept units are grouped into categories, with concepts in the same category more likely to be connected than concepts from different categories. At any time, the set of coactive units in the concept layer encodes an idea. Ideas emerge based on the existing connections between concepts and previous ideas. To lower the probability of an idea to remain active indefinitely or to be expressed again in the near future, connections between coactive units are decreased temporarily.
To test the conceptual network model, we simulated several priming paradigms, similar to those tested in behavioral experiments (Leggett, 1997; Coskun, 2000). For example, we modeled the effect of presenting hints from uncommon or low-accessible categories (i.e., accessed rarely in a particular context) versus hints from common or high-accessible categories (i.e., accessed frequently). Another controlled variable is the number of hints presented (few or many). The semantic network theory predicts that hints from low-accessible categories can access knowledge otherwise not available to a brainstormer, and thus increase the number of ideas and that of original ideas. Results from the conceptual network model confirm this hypothesis and are in line with experimental results. In the model, lowaccessible categories are more difficult to reach, unless primed. Figure 1 shows the activity of excitatory units in all categories when the initial idea is activated in a highaccessible category. It can be seen that bumps of activity are spread mostly over the highaccessible categories (top third of the graph), due to the higher number of incoming connections. Rarely does the activity spread in the middle and lower thirds of the graph, corresponding to medium and low-accessible categories, respectively.
Past research has shown that exposure to the ideas of other group members and/or externally provided hints can have both positive and negative effects on individual performance. Unfortunately, when these other ideas are at odds with an individual’s current train of thought because of conceptual mismatches or inopportune timing – or more likely both – individual performance can be negatively impacted. In order to take advantage of the facilitating effects of the exposure to other ideas, while minimizing the interfering effects, Professor Brown and his students are conducting a series of experiments using a novel paradigm where individuals can request hints at times of their own choosing throughout a brainstorming session. The results so far indicate that individuals requesting hints outperform a control group that is not provided with hints and individuals who choose not to request hints. Preliminary results in Figure 2 show an increasing number of unique ideas as a function of the number of hints requested.
This paradigm allows us to analyze individual responses to specific hints. We are finding some consistent patterns of responses. Individuals often respond to general hints (“We need better parking”) by providing a specific instance (“A parking structure should be built close to the center of campus”). Hints that are more specific (“We need more variety of food in the cafeteria”) tend to result in an expressed modification (“More ethnic cuisine would make the food court less boring”) or a related idea (“More varied types of coffee and tea should be available”).
The generation of novel conceptual combinations is an important source of creativity that has been neglected in group brainstorming research. Some preliminary results have been obtained in another study conducted by Professor Brown, suggesting that individuals may be better than groups at integrating a diverse set of categories into a coherent conceptual theme, but once combined, groups may be superior at coming up with more diverse and more creative instances of the new category (although not necessarily a greater number).
The research by Professors Doboli and Brown is the result of an interdisciplinary collaboration between the fields of psychology and computer science, with the aim of understanding the factors and mechanisms that favor the production of novel and creative ideas in groups as well as in individuals. This work can have implications well beyond brainstorming and idea generation. By uncovering the process by which novel and useful associations in large conceptual spaces are discovered, we can develop artificial systems for inventing novel artifacts, or designing intelligent systems, able to come up with novel responses in unknown situations. Also, more and more common are virtual groups that are distributed across space and time, where communication is done asynchronously via the Internet. It will be very interesting to compare traditional face-to-face groups and virtual groups and to uncover new factors that enhance and inhibit creativity in virtual groups.
This work is supported by collaborative NSF Human and Social Dynamics grants BCS- 0729470 (Simona Doboli, Hofstra University), BCS- 0728413 (Ali A. Minai, University of Cincinnati) and BCS-0729305 (Paul Paulus, University of Texas at Arlington), including funds from the deputy director of National Intelligence for Analysis.
Disclaimer: Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.
Brown, V., Tumeo, M., Larey, T., and Paulus, P.B. (1998). Modeling cognitive interactions during group brainstorming. Small Group Research, vol. 29, pp. 495-526.
DeRosa, D.M., Smith, C.L., and Hantula, D.A. (2007). The medium matters: Mining the long-promised merit of group interaction in creative idea generation tasks in a meta-analysis of the electronic group brainstorming literature. Computers in Human Behavior, vol. 23, pp. 1549-1558.
Diehl, M., and Stroebe, W. (1987). Productivity loss in brainstorming groups: Toward the solution of a riddle. Journal of Personality and Social Psychology, vol. 53, pp. 497-509.
Doboli, S., Minai, A.A., and Best, P.J. (2000). Latent attractors: A model for context-dependent place representations in the hippocampus. Neural Computation, 12: 1009-1043.
Doboli, S., Brown, V., and Minai, A.A. (forthcoming, 2009). A conceptual neural model of idea generation. Proceedings of the International Conference on Neural Networks, Atlanta, June 14-18, 2009.
Doboli, S., Minai, A.A., and Brown, V. (2007). Adaptive dynamic modularity in a connectionist model of context-dependent idea generation. Proceedings of the International Conference on Neural Networks, Orlando, Florida, pp. 2183- 2188.
Iyer, L., Minai, A.A., Doboli, S., and Brown, V. (2007). Modularity and selforganized functional architectures in the brain. Proceedings of the 7th International Conference on Complex Systems, Boston, MA, October 28-November 2, 2007.
Iyer, L., Minai, A.A., Doboli, S., and Brown, V. (2008). Modeling exploration and exploitation in creative idea generation. Proceedings of the 19th Midwest Artificial Intelligence and Cognitive Science Conference, Cincinnati, Ohio, April 12-13, 2008.
Larey, T.S., and Paulus, P.B. (1999), Group preference and convergent tendencies in small groups: A content analysis of group brainstorming performance. Creativity Research Journal , vol. 12, pp. 175-184.
Leggett, K. (1997). The effectiveness of categorical priming in brainstorming. Master’s thesis, University of Texas at Arlington.
Osborn, A. (1957). Applied Imagination. New York: Scribner’s.
Paulus, P.B., and Dzindolet, M. (1993). Social influence processes in group brainstorming. Journal of Personality and Social Psychology, vol. 64, pp. 575-586.
Paulus, P.B., Nakui, T., Putman, V.L., and Brown, V. (2006). Effects of task instructions and brief breaks on brainstorming. Group Dynamics: Theory, Research, and Practice, vol. 10, pp. 206-219.