Something Changed, but Which Thing? Two Scales of Visual Change Detection

You look at thirty red and green disks. After a brief blank interval, the arrangement returns. There seems to be more red, yet you cannot point to a disk that changed. Must this be an unsupported guess? Or can information about a collection differ from information sufficient to identify one of its members? A 2014 study by Piers D. L. Howe and Margaret E. Webb investigated that distinction with concrete visual tasks. [1]

The interesting possibility does not require a mysterious intuition. An ordinary difference in available information may suffice: noticing a change in a group's color proportion is not the same as knowing which members changed. The researchers tested this possibility while estimating observers' tendency to guess. Reporting a change and correctly identifying a changed object were collected as separate responses, rather than treating the second as an automatic consequence of the first. [1]

Starting with portrait differences

In Experiment 1, ten observers saw two portrait photographs of the same woman in succession. Each appeared for 1.5 seconds, separated by a one-second blank. The images did not continuously alternate. This one-shot arrangement was intended to avoid a verification strategy in which an observer first suspects a change and waits for another repetition to confirm it. [1]

Observers first indicated whether a change had occurred. After a yes response, they selected the changed feature from nine options, including earrings, glasses, a hat, and hair. There were one hundred change trials and forty catch trials. None of the nine designated features changed on catch trials, although the person's pose was not held perfectly identical between photographs. Pose differences occurred in the change photographs too. [1]

Consequently, no change did not mean pixel-for-pixel identity. The relevant change was defined by the task. Features were arranged randomly so that looking at only one photograph could not reliably reveal whether a designated item had been added or removed. The observer had to compare the two presentations. This distinction between task-relevant and incidental differences mattered when interpreting a report that something had changed. [1]

Counting a tendency to say yes

Someone can report a change without having useful evidence. Choosing the wrong feature afterward would not, by itself, demonstrate a special perceptual ability. The researchers therefore used false change reports on catch trials to estimate each observer's guessing tendency. Their model predicted how many trials could involve a correct change report followed by incorrect identification because of guessing. [1]

The model was not assumption-free. It distinguished an actual detection from a false report when no change existed, and used no-change trials to estimate responses when a change had not been detected. This was more informative than simply counting incorrect object selections, but the assumptions remained part of the interpretation. The analysis tested whether the observed pattern exceeded what that account of response bias could explain. [1]

Experiment 1 produced more detection-without-identification trials than the guessing model predicted. This supported the possibility that some changes could be reported without enough information to select the correct feature. Experiment 2 repeated the task with inverted portraits and found a similar excess. Inversion did not establish that every special aspect of face processing had disappeared, but the result was not confined to upright portraits. [1]

A proportion changes among the disks

Experiment 3 replaced photographs with a more controllable arrangement of thirty disks. Each disk was independently assigned red or green at random. On change trials, three disks of the same color switched to the other color, altering the overall proportion of red and green. Observers reported whether a change occurred and then selected one disk they thought had changed. Selecting any of the three changed disks counted as correct; identifying all three was not required. [1]

The arrangement made the distinction between overall appearance and object-specific knowledge particularly clear. Red could occupy a larger share without the observer knowing which three locations produced the increase. Because the initial colors were random, a single image with plenty of red was not enough to establish a change. The useful information had to concern the difference between the two presentations. [1]

Detection followed by incorrect disk selection exceeded the guessing prediction substantially. Each of the ten observers showed more such trials than the model expected. This did not mean everyone detected every change, or that an incorrect selection established a complete absence of partial location information. The narrower finding was that reports on some trials lacking correct object identification could not adequately be explained as random yes responses. [1]

Keeping the color amounts constant

Experiment 4 used fifteen rectangles, each half red and half green. On change trials, the colored halves of two rectangles switched places. The overall amount of red and green remained constant, removing the usefulness of the color-proportion cue in its previous form. Observers again reported a change before attempting object selection. [1]

This time, detection-without-identification trials were not significantly more frequent than the guessing model predicted. A direct comparison between Experiments 3 and 4 also showed a difference. Together these results supported the color-proportion account. However, the displays differed in other respects too, including geometry and object number. It would be too strong to describe this as isolating one variable while holding every other feature constant. [1]

The control also weakened explanations based solely on accidental clicks or immediately forgetting which object changed. Such explanations would otherwise have been expected to produce the same broad difficulty in the control task. Nevertheless, a nonsignificant excess is not proof that detection without identification is absolutely impossible in that arrangement. Evidence from small samples in particular tasks supports a more bounded statement than a universal limit on perception.

Where partial knowledge is enough

The authors explicitly noted that they had not measured how precisely observers could approximately localize an unidentified change. The experiments also did not settle conscious awareness of the change. Failing to select an exact object does not show that someone had no information about it whatsoever. A person might know that the display became redder without knowing which disk supplied that evidence. [1]

There is therefore no need to recast the finding as a sixth sense or a complete hidden record of every detail. The more useful conclusion concerns levels of information. Knowledge of an overall change need not support the correct identity of an individual member. The study's contribution was to make that distinction testable through catch trials, a guessing account, and a change in available cues. Knowing that something changed and finding the thing that changed are different measures of success.

Sources

[1] Howe, P. D. L., and Webb, M. E. (2014). Detecting Unidentified Changes. PLOS ONE, 9(1), e84490. https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0084490 . DOI: 10.1371/journal.pone.0084490. Independently written explanation of research published under a Creative Commons Attribution license.