Informing the design of computer-based environments to support creativity

Introduction According to legend, Archimedes shouted “Eureka!” (“I have found it!”) when he suddenly discovered that water displacement could be used to calculate density. Since then, “Eureka!,” or “Aha!,” has often been used to express the feeling one gets when solving a problem with insight. Insight is pervasive in human (and possibly animal [Epstein et al. 1984]) cognition, occurring in perception, memory retrieval, language comprehension, problem solving, and various forms of practical, artistic, and scientific creativity (Sternberg and Davidson 1995). The Archimedes legend has persisted over two millennia in part because it illustrates some of the key ways in which insight solutions differ from solutions achieved through more straightforward problem solving. We examine the neural bases of these different problem-solving methods. Although many processes are shared by most types of problem solving, insight solutions appear to differ from noninsight solutions in several important ways. The clearest defining characteristic of insight problem solving is the subjective “Aha!” or “Eureka!” experience that follows insight solutions (Schooler et al. 1993). This subjective experience can lead to a strong emotional response—according to legend, Archimedes ran home from the baths shouting “Eureka!” without donning his clothes first. In addition, problem solving with insight is characterized by the following features. (1) Solvers first come to an impasse, no longer progressing toward a solution (Duncker 1945). Archimedes, for example, was stymied by King Hiero's challenge to determine whether his new crown was pure gold without damaging the crown. (2) Solvers usually cannot report the processing that enables them to reinterpret the problem and overcome the impasse (Maier 1931). Insight often occurs when people are not even aware they are thinking of the problem, as reportedly happened to Archimedes while in the baths. (3) Solvers experience their solutions as arising suddenly (Metcalfe and Wiebe 1987; Smith and Kounios 1996) and immediately recognize the correctness of the solution (or solution path). (4) Performance on insight problems is associated with creative thinking and other cognitive abilities different from those associated with performance on noninsight problems (Schooler and Melcher 1997). Some researchers have argued that all these characteristics of insight solutions are essentially epiphenomenal, that insight and noninsight solutions vary only in emotional intensity, and that they are attained with precisely the same cognitive (hence neural) mechanisms (Weisberg and Alba 1981; Weisberg 1986; Perkins 2000). Persistent questions about insight concern whether unconscious processing precedes reinterpretation and solution, whether distinct cognitive and neural mechanisms beyond a common problem-solving network are involved in insight, and whether the apparent suddenness of insight solutions reflects truly sudden changes in cognitive processing and neural activity. Recent work suggests that people are thinking—at an unconscious level—about the solution prior to solving problems with insight. Specifically, while working on a verbal problem they have yet to solve, people presented with a potential solution word read the actual solution word faster than they read an unrelated word (Bowden and Beeman 1998). This “solution priming” effect is greater—and in fact people make solution decisions about presented words more quickly—when words are presented to the left visual hemifield, which projects directly to the right hemisphere (RH), than when words are presented to the right visual hemifield, which projects to the left hemisphere (LH). This suggests that RH semantic processing is more likely than LH semantic processing to produce lexical or semantic information that leads to the solution. These RH advantages occur only when solvers experience insight—the “Aha!” or “Eureka!” feeling that comes with insight solutions (Bowden and Jung-Beeman 2003a). Moreover, when subjects try to solve classic insight problems, they benefit more from hints presented to the left visual field (i.e., the RH) than from hints presented to the right visual field (i.e., the LH) (Fiore and Schooler 1998). Problem solving is a complex behavior that requires a network of cortical areas for all types of solving strategies and solutions, so solving problems with and without insight likely invokes many shared cognitive processes and neural mechanisms. One critical cognitive process distinguishing insight solutions from noninsight solutions is that solving with insight requires solvers to recognize distant or novel semantic (or associative) relations; hence, insight-specific neural activity should reflect that process. The most likely area to contribute to this component of insight problem solving is the anterior superior temporal gyrus (aSTG) of the RH. Language comprehension studies demonstrate that the RH is particularly important for recognizing distant semantic relations (Chiarello et al. 1990; Beeman 1998), and bilateral aSTG is involved in semantic integration. For example, sentences and complex discourse increase neural activity in aSTG bilaterally (Mazoyer et al. 1993; Stowe et al. 1999), and discourse that places particular demands on recognizing or computing distant semantic relations specifically increases neural activity in RH temporal areas (St. George et al. 1999; Mason and Just 2004), especially aSTG (Meyer et al. 2000; Kircher et al. 2001). If this prediction of RH aSTG involvement is confirmed, it will help constrain neurocognitive theories of insight. Other cortical areas, such as prefrontal cortex and the anterior cingulate (AC) may also be differentially involved in producing insight and noninsight solutions. We used functional magnetic resonance imaging (FMRI) in Experiment 1 and electroencephalogram (EEG) measurement in Experiment 2 to test the empirically and theoretically derived hypothesis that solving problems with insight requires engagement of (or increased emphasis on) distinct neural mechanisms, particularly in the RH anterior temporal lobe. Event-related experimental designs compared neural activity when people solved verbal problems with insight to neural activity when they solved problems (from the same problem set) without insight. As in earlier behavioral work, we used a set of compound remote associate problems (Bowden and Jung-Beeman 2003b) adapted from a test of creative cognition (Mednick 1962). Figure 1 illustrates the sequence for each trial. Subjects saw three problem words (pine, crab, sauce) and attempted to produce a single solution word (apple) that can form a familiar compound word or phrase with each of the three problem words (pineapple, crab apple, applesauce). We relied on solvers' reports to sort solutions into insight solutions and noninsight solutions, avoiding the complication that presumed insight problems can sometimes be solved without insight (Davidson 1995) and circumventing the use of different types of problems requiring different cognitive operations. Thus, we made use of the most important defining characteristic of insight problems: the subjective conscious experience—the “Aha!” A similar technique revealed distinct behavioral characteristics when people recognized solutions with insight (Bowden and Jung-Beeman 2003a). Note that this is a very “tight” comparison. In both conditions problems are solved using a network of processes common to both insight and noninsight solutions. If insight ratings reflect some distinct cognitive processes, this contrast will reveal the distinct underlying brain activity. In other words, within the cortical network for problem solving, different components will be engaged or emphasized for insight versus noninsight solutions. FMRI (Experiment 1) should reveal neuroanatomical locations of processes that are unique to insight solutions, and EEG (Experiment 2) should reveal the time course (e.g., whether insight really is sudden) and frequency characteristics of neurophysiological differences. Results Experiment 1 Subjects solved 59% of the problems presented, and pressed buttons indicating “insight” for 56% (s.d. = 18.2) of their solutions, “no insight” for 41% (s.d. = 18.9) of their solutions, and “other” for 2% of their solutions. We marked a point about 2 s (rounded to the nearest whole second) prior to each solution button press as the solution event, and examined a time window 4–9 s after this event (i.e., 2–7 s after the button press) to isolate the corresponding hemodynamic response. Solving problems and responding to them required a strict sequence of events (reading of words, solving effort, solving, button press, verbalizing the solution, insight decision), but this sequence was identical whether subjects indicated solving with or without insight, so differences in FMRI signal resulted from the degree to which distinct cognitive processes and neural systems led to insight or noninsight solutions. Figure 2 illustrates the most robust insight effect: as predicted, insight solutions were associated with greater neural activity in the RH aSTG than noninsight solutions. The active area was slightly anterior to primary auditory cortex, posterior to temporal pole, and along the medial aspect of the aSTG, extending down the lateral edge of the descending ramus of the Sylvian fissure to midway through the middle temporal gyrus (MTG). (This site is also close to the superior temporal sulcus, which has been implicated in language). Across all 13 subjects, the peak signal difference at a single voxel within the RH aSTG was 0.25% across the 6-s window, and 0.30% at a single time to repetition (TR), i.e., the time needed to repeat the image of the whole brain. Overall signal in this region was robust, reaching 96.8% of the brainwide average (after removing voxels in other brain areas with signal below a standard criterion). Within the cluster of voxels identified across the group, 12 subjects showed from 0.03% to 0.35% greater signal for insight than for noninsight solutions; one subject showed 0.02% greater signal for the noninsight solutions. It is not likely that RH aSTG is involved only in output or in emotional response following insight solutions, because neural activity in this area also increased when subjects first encountered each problem (Figure 3). Thus, RH aSTG is involved in processing the problem words both initially and at solution. (Of course, event-related FMRI signal occurred in many other cortical regions at problem onset, especially visual cortex). There was no insight effect in response windows immediately preceding or following the defined response window. All indications point to a striking transient event in the RH aSTG near the time when subjects solve problems with insight. The involvement of the RH rather than the LH for this verbal task is not due to greater difficulty in producing insight solutions: subjects produced insight solutions at least as quickly (mean solution time = 10.25 s, s.d. = 3.58 s) as they produced noninsight solutions (mean = 11.28 s, s.d. = 4.13 s) (t 0.3). More importantly, the hemodynamic responses to both insight and noninsight solutions in the homologous area of the LH are about equivalent to the response to noninsight solutions in the RH aSTG—it is the strong response to insight solutions in the RH aSTG that stands out. There is no insight effect anywhere within temporal cortex of the LH. At statistical thresholds below significant levels (p 3.43, p < 0.005 uncorrected). Monte Carlo simulations with similar datasets reveal low false positive rates with these criteria. RH aSTG was the only cluster to exceed these criteria, and converging evidence and the a priori prediction about RH aSTG strengthen confidence in this result. Experiment 2 Behavioral procedures were similar to those of Experiment 1, except that (A) problem words were presented at smaller visual angles to discourage eye movements, (B) there were 2-s delays between each event in the response sequence, and (C) subjects triggered a new problem directly after responding to the previous problem (i.e., no line task occurred between problems). EEG methods Continuous high-density EEGs were recorded at 250 Hz (bandpass: 0.2–100 Hz) from 128 tin electrodes embedded in an elastic cap (linked mastoid reference with forehead ground) placed according to the extended International 10–20 System. Prior to data analysis, EEG channels with excessive noise were replaced with interpolated data from neighboring channels. Eyeblink artifacts were removed from the EEG with an adaptive filter separately constructed for each subject using EMSE 5.0 (Source Signal Imaging Inc., San Diego, California, United States). Induced oscillations were analyzed by segmenting each subject's continuous EEG into 4-s segments beginning 3 s before each solution response. (An analysis epoch beginning at an earlier point in time would have resulted in the loss of trials associated with response times of less than 3 s.) Time-frequency transforms (performed with EMSE 5.0) were obtained by the application of complex-valued Grossmann-Morlet wavelets, which are Gaussian in both time and frequency. Following Torrence and Campo (1998), the mother wavelet, ω0, in the time domain has the form where ω0 is a nondimensional frequency. In this case, ω0 is chosen to be 5.336, so that ∫ϕ0(t) ≅ 0. The constant π−¼ is a normalization factor such that ∫(ϕ0(t))2 = 1. For the discrete time case, a family of wavelets may be obtained as where δt is the sample period (in seconds), s is the scale (in seconds), and n is an integer that counts the number of samples from the starting time. The Fourier wavelength λ is given by In the frequency domain, the (continuous) Fourier transform of Equation 2 is where One reasonable way to measure the “resolution” of the wavelet transform is to consider the dispersion of the wavelets in both time and frequency. Since the wavelets are Gaussian in both domains, the e-folding time and frequency may serve as quantitative measures of dispersion. Note that these dispersions are a function of the scale, s. For a selected frequency, 𝒻 c = 1/λ, or from Equation 3 Then substituting into Equation 2, we find that the e-folding time is for frequency 𝒻 c . From Equation 2, the e-folding frequency is . To make this concrete, we find that for a 10-Hz (alpha-band) center frequency, the e-folding time is 0.12 s and the e-folding frequency is 2.6 Hz. For a 40-Hz ( gamma-band) center frequency, the e-folding time is 0.03 s and the e-folding frequency is 10.5 Hz. Note that these e-folding parameters imply that wavelet scaling preserves the joint time-frequency resolution (equal areas in time-frequency space), with higher temporal resolution but broader frequency resolution as the wavelet scale decreases. Segments corresponding to trials for which individual subjects produced the correct response were isolated and averaged separately according to whether or not the subject reported the experience of insight. Planned statistical tests (repeated-measure ANOVAs) were performed in order to detect insight-related effects on frontal midline theta (5–8 Hz), posterior alpha (8–13 Hz), fronto-central beta (13–20 Hz), and left and right temporal gamma (20–50 Hz). Response-locked event-related potentials (ERPs) were also computed using the same analysis epoch. Standard ERP analyses yielded no evidence of statistically significant effects, likely because ERPs reflect phase-locked activity rather than the induced (i.e., nonphase-locked) activity examined in the wavelet analyses; due to the long response times evident in this experiment, phase locking resulting from problem presentation would not be expected. EEG effects were topographically mapped by employing spline-based Laplacian mapping with an FMRI-derived realistic head model and digitized electrode positions. Localization of EEG/ERP signals is a form of probabilistic modelling rather than direct neuroimaging. In contrast to other techniques, source estimation by Laplacian mapping indicates the presence of superficial foci of neuroelectric activity with minimal assumptions. Supporting Information Figure S1 Cortical Regions Showing “Insight Effects” Below Cluster Size Threshold The far left lane shows for each region a single slice best depicting the cluster activated above threshold; middle lane shows time course of signal following insight (red line) and noninsight (blue line) solutions, across the entire active cluster; right panel shows the “insight effect” (insight signal minus noninsight signal, error bars show the standard error of the mean of the difference at each timepoint). (A) depicts bilateral IFG with lowered threshold (t[12] = 2.83, p < 0.015); (B–D) depict clusters of FMRI signal at the same t-threshold used in the main paper (t[12] = 3.43, p < 0.005), but the clusters are too small to surpass cluster criterion. (B) LH medial frontal gyrus; (C) LH PC gyrus; (D) LH amygdala (there was also a small cluster near RH amygdala). Spatial coordinates and other are details listed in Table 1. (914 KB PDF). Click here for additional data file.