There are three on-off light switches on the wall of the first floor of a building. One of the switches is initially off and controls an incandescent bulb in a lamp on the third floor of the building. The other two switches do not control the bulb or anything else (they are disconnected). How can you find out which one of the three switches turns the light bulb on and off? You can toggle the switches as many times as you want and for as long as you want, but you can walk only once to the third floor to check on the light bulb.
While you’re working on that, I’ll show you what you look like (no offense!):
(A hint and the solution to the problem are at the bottom of this post.)
It’s thrilling when it happens, but what actually causes insight? New research in the Journal of Cognitive Neuroscience takes us one step closer to an answer: up to 8 seconds before people solve problems thought to require insight, a particular set of very fast oscillations are observable above the right frontal lobe.
Sheth, Sandkuehler & Bhattacharya gave 18 subjects a series of “insight problems” like the one at the start of this post, while the electrical activity on subjects’ scalp was recorded via a sensor net with 32 electrodes. All the problems shared a number of features:
- no problem required specialized knowledge, nor could any be solved with a predefined procedure
- if the correct solution was described, it would seem obvious (i.e., they were simple puzzles)
- the puzzles weren’t well known, and didn’t require pen/paper to solve
To get the gist of this, contrast these criteria with algebra problems (which do require specialized knowledge, can be solved with predefined procedures, often with the help of pen and paper, and do not have typically yield answers that are transparently correct). If subjects hadnt’ solved a problem within 60-90s, they were given a hint. After an additional 60-90s, or until subjects felt they’d solved the problem, subjects first indicated whether they felt a sensation of insight (on a scale from 0-10), and then verbally described their solution. Data from the sensor net were transformed with Morlet wavelets to quantify the power of oscillations at particular frequencies on the scalp, and then subjected to something like principal components analysis (a technique the authors call “PARAFAC,” a form of factor analysis used to identify the “scalp topography, spectral frequency, and temporal dynamics that provide maximum discriminatory power between two compared conditions.”)
The results showed two primary effects discriminating correct from incorrect solutions: a decrease in the power of frequencies between 15-25 Hz (the so-called “beta-band”, sometimes implicated in active maintenance) over posterior regions, and an increase in the power of frequencies between 30-75 Hz (the so-called “gamma-band”, implicated in higher cognitive functioning) over right frontal regions. The beta-band reduction was more prominent in:
1) solved problems relative to unsolved problems;
2) solved problems where no hint had been provided relative to those with hints;
3) solved problems with a hint relative to those remaining unsolved despite the hint,
4) in problems solved with a strong feeling of insight relative to those without this feeling.
Similar results were found for increases in the gamma-band, in every case except for #3, in which gamma band increases were actually more pronounced for problems remaining unsolved despite the hint. The authors suggest this could reflect that the “transformative thought” underlying insight is reduced by hints – particularly when those hints lead to a solution. Perhaps subjects are engaging in a larger number of transformative thought processes (just the wrong ones) following a hint that does not ultimately yield a successful solution.
Many of these oscillatory changes preceded subjects’ responses by several seconds (up to 8 seconds, in the case of the gamma-band increases), leading to the suggestion that this reflects unconscious processing (as discussed at the Economist). i think the flaw in this reasoning is made clear with a simple example.
Let’s suppose that some subjects were allowed to close their eyes instead of reading the problems. Clearly we would be able to determine which subjects were most likely to solve the problems long before any actually solved them, likely by looking at activity in the visual cortex. However, we would not say that this activity in the visual cortex reflects unconscious processing. The best evidence we have that insight problems reflect “unconscious” processing is that people have difficulty reporting how they solve them – in other words, the brain data add little to this debate.
There are a number of other caveats to the research as well. First, no one really knows what processes any of these oscillations reflect, nor do we know their sources. Second, the changes in oscillatory power are inherently ambiguous, because they can reflect many underlying changes in the brain: an change in neuronal firing rates, a change in the coherence with which neurons are firing (without changes in rate), or a change in the phase coherence across multiple frequencies with which neurons are firing. Third, and most critically, Sheth et al used an unusual baseline condition in calculating the changes in oscillatory power: the time during which subjects were reading the problems.
The problem with this baseline can be easily illustrated. Let’s suppose that subjects who ultimately solve the problems undergo more “transformational thought” while they’re first reading the problem. This seems like a reasonable assumption, given that one’s understanding of the problem is likely to be restructured as you integrate the various constraints, and that to the extent this occurs one is more likely to actually solve the problem. Thus, these “better subjects” will show less beta-band power during their (mostly correct) solutions (relative to baseline) than the “worse subjects” will show during their (mostly incorrect or undiscovered) solutions (relative to baseline). Whether this particular example applies in the case of Sheth et al is hard to say, but it illustrates the general problem of using such a cognitively “high-level” baseline, and highlights the underlying ambiguity with what these oscillations actually mean.
If you still haven’t solved the problem, here’s your hint: Turn one switch on for an hour and then turn it off. The solution is in the comments.