Complex problem solving requires a complex cortical network to encode the problem information, search memory for relevant information, evaluate this information, apply operators, and so forth. The FMRI and EEG results reported here conclusively demonstrate that solving verbal problems with insight requires at least one additional component to this cortical network, involving RH aSTG, that is less important to solving without insight. The insight effect in RH aSTG accords with the literature on integrating distant or novel semantic relations during language comprehension. When people comprehend (read or listen to) sentences or stories, neural activity increases in aSTG or temporal pole bilaterally more than when comprehending single words (Mazoyer et al. 1993
; Bottini et al. 1994
; Stowe et al. 1999
; Humphries et al. 2001
; Meyer et al. 2000
). Neural activity increases in predominantly RH aSTG during tasks that emphasize integration across sentences to extract themes (St. George et al. 1999
) or to form more coherent memories for stories (Mason and Just 2004
). RH aSTG is also selectively active when subjects must generate the best ending to a sentence (Kircher et al. 2001
) or mentally repair grammatically incorrect sentences (Meyer et al. 2000
), both of which likely require intense semantic integration.
Like the results in language processing, the current results are predicted by the theory that the RH performs relatively coarse semantic coding (Beeman 1998
; similarly, Chiarello et al. 1990
). This theory contends that when people encounter words, semantic processing in several LH areas engages in relatively fine semantic coding which produces small semantic fields—i.e., this processing strongly focuses on a few concepts closely related to the input word in the given context. This is very effective for most straightforward language processing. In contrast, the homologous RH areas engage in relatively coarse semantic coding, which produces large and weak semantic fields—i.e., this processing includes many concepts, even concepts distantly related to the input words and context. This process is ineffective for rapid interpretation or selection but increases semantic overlap among multiple semantic fields (Beeman et al.
1994), which is useful when drawing together parts of a story or conversation that are only distantly related (Beeman 1993
; Beeman et al. 2000
). In this view, the coarseness of semantic coding is largely influenced by slight asymmetries in neural microcircuitry that produce more discrete, less redundant input fields in pyramidal neurons of the LH language cortex, and more overlapping input fields in corresponding neurons in the RH (for reviews see Beeman 1998
; Hutsler and Galuske 2003
We suggest that semantic integration, generally, is important for connecting various problem elements together and connecting the problem to the solution, and that coarsely coded semantic integration, computed in RH aSTG, is especially critical to insight solutions, at least for verbal problems (or problems that can be solved with verbal or semantic information). People come to an impasse on insight problems because their retrieval efforts are misdirected by ambiguous information in the problem or by their usual method for solving similar problems. Large semantic fields allowing for more overlap among distantly related concepts (or distantly associated lexical items) may help overcome this impasse. Because this semantic processing is weak, it may remain unconscious, perhaps overshadowed by stronger processing of the misdirected information (Schooler et al. 1993
; Smith 1995
), and solvers remain stuck at impasse. Eventually, solution-related information bursts into awareness “in a sudden flash.” This can happen after misdirected processing decays or is suppressed, after solution-related processing grows, or after environmental cues occur—such as the water overflowing the bathtub when Archimedes got in. Archimedes had semantic and verbal knowledge about how to compute density from weight and volume, but struggled with measuring the volume of an irregularly shaped crown without harming the crown (e.g., melting it). His observation of water displacement allowed him to connect known concepts in new ways. This is the nature of many insights, the recognition of new connections across existing knowledge.
A persistent question has been whether the cognitive and neural events that lead to insight are as sudden as the subjective experience. The timing and frequency characteristics of the EEG results shed light on this question. We propose that the gamma-band insight effect in Experiment 2
reflects the sudden transition of solution-related cognitive processing from an unconscious to a conscious state. Recent research associates gamma-band oscillations with the ignition of neural cell assemblies supporting the transient feature binding necessary to activate a representation (Tallon-Baudry and Bertrand 1999
; Pulvermüller 2001
)—in this case, a phonological, lexical, or semantic representation corresponding to the solution word and its associations to the problem words. According to this hypothesis, greater synchronous gamma-band activity for insight than for noninsight solutions could reflect a more integrated or unitized solution representation. Furthermore, synchronous gamma-band activity has been hypothesized to play a critical role in the accessibility to consciousness of such representations (Engel and Singer 2001
). The timing (with respect to the solution button press) of the insight gamma-band effect closely approximates estimates derived from cognitive behavioral studies of the amount of time required to access an available solution and generate a two-alternative, forced-choice button-press response (e.g., Kounios et al. 1987
; Meyer et al. 1988
; Smith and Kounios 1996
). The present experiments had no response choice (i.e., always the same bimanual button press for solutions), so subjects could easily have responded 0.3 s after solving the problems. Thus, we infer that the observed gamma burst reflects the sudden conscious availability of a solution word resulting from an insight.
Suddenly recognizing new connections between problem elements is a hallmark of insight, but it is only one component of a large cortical network necessary for solving problems with insight, and recognizing new connections likely contributes to other tasks, such as understanding metaphors (Bottini et al. 1994
) and deriving a story theme (St. George et al. 1999
). Similar tasks may depend on related cortical networks. For example, appreciating semantic jokes (Goel and Dolan 2001
) and engaging in deductive reasoning that sometimes involves insight (Parsons and Osherson 2001
) both increase activity in RH posterior MTG. It is striking that the insight effect observed in the RH in our experiments occurred when people solved verbal problems, which traditional views suggest should involve mostly LH processing with little or no contribution from the RH. It is possible that insight solutions to nonverbal problems would require different cortical networks. However, the observed effect cannot be due simply to verbal retrieval, which must occur for both insight and noninsight solutions; it could be due to a type of verbal retrieval specific to insight solutions, but not involved in noninsight solutions.
We turn now to another result from the EEG time-frequency analysis, which was not predicted but nevertheless suggests a provocative interpretation. The gamma burst thought to reflect the transition of the insight solution from an unconscious to a conscious state was preceded by insight-specific activity in the alpha band (8–13 Hz). Specifically, there was a burst of alpha power (estimated at 9.8 Hz) associated with insight solutions detected over right posterior parietal cortex from approximately 1.4 s until approximately 0.4 s before the solution response, at which point insight alpha power decreased to the level of noninsight alpha power, or below (). An ANOVA was performed on log-transformed alpha-band (9.8 Hz) EEG power at left and right parietal-occipital electrode sites (PO7 and PO8, respectively) for insight and noninsight trials using three time windows: −2.06 to −1.56 s, −1.31 to −0.56 s, and −0.31 to 0.06 s (measured from the solution button press). This analysis yielded a significant insight × time window interaction (F[2,36] = 4.13, p = 0.027, with the Huynh-Feldt correction). Follow-up t-tests in each time window yielded significant effects of insight in the first time window at both electrode sites (PO7: t = 2.32, p = 0.033; PO8: t = 2.42, p = 0.026) and in the second time window only at the RH site (PO8: t = 2.17, p = 0.043), with a reversal of the direction of the effect. The third time window yielded no significant effects.
Alpha-Band Power for Insight and Noninsight Solutions
Alpha rhythms are understood to reflect idling or inhibition of cortical areas (Pfurtscheller et al. 1996
). Increased alpha power measured over parietal-occipital cortex indicates idling or inhibition of visual cortex. This has been attributed to gating of visual information flowing into the perceptual system in order to protect fragile or resource-intensive processes from interference from bottom-up stimulation (Ray and Cole 1985
; Worden et al. 2001
; Jensen et al. 2002
; Cooper et al. 2003
; Ward 2003
). This interpretation assumes that brain areas are normally highly interactive, and that allowing one process to proceed relatively independently requires active attenuation of this interaction. For instance, when subjects attend to visual space in the hemifield projecting to one hemisphere, posterior alpha increases over the other hemisphere, which receives inputs from the unattended hemifield (Worden et al. 2001
). Analogously, the present results suggest selective gating of visual inputs to the RH during the interval preceding the insight-related right temporal gamma burst (). Hypothetically, this allows weaker processing about more distant associations between the problem words and potential solutions to gain strength, by attenuating bottom-up activation or other neural activity not related to solution that would decrease the signal-to-noise ratio for the actual solution.
The Time Course of the Insight Effect
This interpretation of the early insight-specific alpha effect is consistent with previous behavioral research suggesting that, prior to an insight, the solution to a verbal problem can be weakly activated (Bowers et al. 1990
), especially in the RH (Bowden and Beeman 1998
; Bowden and Jung-Beeman 2003a
). Thus insight solutions may be associated with early unconscious solution-related processing, followed by a sudden transition to full awareness of the solution. We suggest that, in Experiment 2
, the early posterior alpha insight effect is an indirect correlate of the former, and the right temporal gamma effect is a direct correlate of the latter.
In sum, when people solve problems with insight, leading to an “Aha!” experience, their solutions are accompanied by a striking increase in neural activity in RH aSTG. Thus, within the network of cortical areas required for problem solving, different components are engaged or emphasized when solving with versus without insight. We propose that the RH aSTG facilitates integration of information across distant lexical or semantic relations, allowing solvers to see connections that had previously eluded them. In the two millennia since Archimedes shouted “Eureka!,” it has seemed common knowledge that people sometimes solve problems—whether great scientific questions or trivial puzzles—by a seemingly distinct mechanism called insight. This mechanism involves suddenly seeing a problem in a new light, often without awareness of how that new light was switched on. We have demonstrated that insight solutions are indeed associated with a discrete, distinct pattern of neural activity, supporting unique cognitive processes.