Doboli, S., & Brown, V. (2006, May) A Neural Network Simulation of Interactive Group Brainstorming. Eighteenth Annual Convention of the Association for Psychological Science. New York.

Groups and teams are popularly considered an effective means of generating creative ideas (e.g. in the field of advertising), and they are widely used. Interestingly, however, in controlled laboratory settings interactive groups nearly always generate fewer ideas than an equal number of individuals brainstorming alone.

Although a great deal of research has been directed towards understanding this "group productivity gap," until recently there have been no attempts to formally model the cognitive processes that underlie individual and group idea generation. One such effort is the associative memory model of Brown and Paulus (2002)*, which assumes that ideas are generated in a process similar to spreading activation in a semantic network. As useful as this model has been, it has some important limitations. The model's relatively high level of abstraction provides very few hints about underlying mechanics (e.g., the neural implementation) of the brainstorming process. The current version of the model is implemented at a level of abstraction that does not represent individual ideas, but only categories. Since the model assumes simply that ideas are retrieved from existing knowledge stored in long-term memory, there is no mechanism for creating novel combinations of ideas. Novel combinations are often considered to be an important aspect of creativity. Although there may be a number of ways to extend the model to address these shortcomings, a connectionist, or neural-network, representation has a number of potential advantages. In addition to providing a closer analogy to the underlying physiological processes, a distributed network allows for the possibility that existing idea representations (attractors in the dynamics of the network) can be combined in novel formations. This presentation describes a hierarchical neural network model of interactive group brainstorming. The long-term memory of a group member is modeled as a single-layer recurrent network. Categories are stored as attractors in the network. A number of individual ideas are encoded in each category. New ideas correspond to a combination of activities not previously stored in the network. Beside recurrent connections, the network also has external inputs. The external input has a random noise component and -- when brainstorming in a group -- a component representing the ideas expressed by other group members. The activation threshold of each neuron is dynamic: whenever a neuron fires, its activation threshold increases. This represents the reduced probability of activating the same idea more than once.

Model simulations demonstrate that the neural network extends the earlier model of Brown and Paulus. When compared to an equal number of simulated individuals, simulated interactive groups of four generate more total ideas, more ideas from less-frequently-thought-of (i.e., less accessible) categories, and more novel combinations of existing ideas. Thus, although groups seem to be less effective for brainstorming than individuals in experimental settings, the model simulations suggest that there may be advantages to generating ideas in a group, especially if sheer quantity of ideas is not the only goal.

*Making group brainstorming more effective: Recommendations from an associative memory perspective. Current Directions in Psychological Science, 11, 208-212.


Levine, D., Brown, V., & Gorfein, D. (2006, May) Is it Sun or Son? A Neural Network Implementation of the Activation-Selection Model of Meaning Selection. Tenth International Conference on Cognitive and Neural Systems. Boston.


Brown, V., Gorfein, D., & Amster, H. (2005, May) Order Effects in the Semantic Priming of Homographs: An Application of the Activation Selection Model. Paper presented at the Seventy-Seventh Annual Meeting of the Midwestern Psychological Association. Chicago.


Cox, B.D., Wandas, T., & Brown, V. R. (2005, April) A Multinomial Source-Monitoring Analysis of Preschoolers Ability to Distinguish Described from Experienced Events.  Poster presented at the 2005 Biennial Meeting of the Society for Research in Child Development. Atlanta.
 
  This poster describes a recoding and reanalysis of a study of the effects of nonsuggestive memory enhancement and of interview preparation (“ground rules” instruction) on preschoolers’ source monitoring ability (Wandas & Cox, 2003).  This type of  “misinformation” study has often been used as a laboratory analog of the suggestive effects of repeated interviewing.  Following

Poole and Lindsay (1995), 76 preschoolers participated in two science demonstrations and were then asked to describe each demonstration.  Children were then randomly assigned to one of four interviewing protocols:  memory enhancement (ME), interview preparation  (IP), memory enhancement plus interview preparation (MEIP), and a control group which was subject to neither protocol. In the memory enhancement groups (ME and MEIP), the interviewer repeated verbatim every element in each child's recall, and then repeated the child’s entire report verbatim. Ten days later, parents read their children descriptions of two science demonstrations on successive days that included one event the child had originally experienced and one event not previously experienced.  On day 14, children in the interview protocol groups (IP and MEIP) were also subjected to a truth-and-lie ceremony, training to say "I don't know" where appropriate, and a practice interview.  Children’s source monitoring ability can be assessed directly by asking whether a given piece of information was 1) seen in the original event and 2) whether it was in the story.  Pairs of such questions, with each child receiving only two of four demonstrations, produces four states: experienced, heard about, both experienced and heard, or neither.
    The multinomial source-monitoring model described by Batchelder & Riefer (1990) and Riefer, Hu, and Batchelder (1994) provides a technique for evaluating whether children can discriminate old from new information and subsequently identify the source of the information.  It also provides estimates of guessing bias.  Although this approach has been used on source monitoring data before, this analysis is its first use with the 4-state misinformation paradigm.  The model provides a coherent, integrated framework for examining the 4 (source) x 4 (response) x 4 (interview protocol) array of data obtained in this study.
    In general, an examination of the model parameters shows that children recognize scenarios that they actually experienced as “old” but have difficulty determining whether they have merely heard something or not.  Those who did identify story information were fairly accurate at identifying it as something which was read to them as opposed to something they had experienced, but when children only experienced a scenario, they were likely to say they heard it as well.  Children also exhibited a tendency to guess that they had both experienced and been read a scenario when unsure of its source.
    The analysis also indicates that children provided with memory enhancement were better at identifying the source of a scenario (read, experienced, or both) than children that were not provided with memory enhancement.  There was also a tendency for the interview preparation to reduce bias towards responding “both.”


Brown, V. & Gorfein, D. (2004, November) Another Look at Recognition Buildup in the Distractor Paradigm: PI Build-Up and Release Mediated by Operation Span.  Poster presented at the Forty-Fifth Annual Meeting of the Psychonomic Society. Minneapolis.

The activation-selection model of determining the meaning of an ambiguous word or phrase (Gorfein, 2001) is unique in that it does not require an explicit mechanism for suppressing the representation of the nonselected meaning. The model assumes that a meaning is selected when a threshold number of attributes associated with that particular meaning are activated. The interpretation of ambiguous words is directed by the "set principle" which states that processing an ambiguous word in the context of currently active attributes will influence the meaning of the word towards the meaning most strongly associated with the currently active attributes. This provides a mechanism by which meaning selection at one time can affect meaning selection at a later time. Computer simulations of the activation-selection model demonstrate the viability of the approach in helping understand both short- and long-term effects of prior meaning selection on the selection of meaning in a new context.