Developing Intelligence

Say you are writing an email when the phone rings. After the phone call, you return to finish the email. Are you slower to continue writing this email than you would be if you’d been doing something else prior to the phone call? In general, yes – at least according to the finding known as lag-2 repetition cost.

This idea has been tested in an experimental framework by having people perform three tasks (A, B and C) on the same set of stimuli. The critical question is whether you’re slower to complete Task A if the previous trial order was A-B than if it was C-B. In general, you are slower on the current task if it is a repeat of the task performed 2 trials ago – this slowing is known as “lag-2 repetition cost.”

The most common explanation of this finding is that you actively inhibit task A in order to switch to task B – and that this inhibition needs to be undone when you need to switch back to task A, resulting in a reaction time cost relative to situations where A had not been so recently performed. This is known as “backward inhibition.”

This backward inhibition hypothesis fails to straightforwardly account for new data reported by Druey & Hubner in the current issue of Psychonomic Bulletin and Review. The authors realized that studies showing lag-2 repetition cost usually employed a particular design where the task cues (the icons that indicate which task subjects should perform) were always visible at the same time as the stimuli requiring a response. The authors went on to demonstrate that no lag-2 repetition costs are reliably observed when task cues and stimuli are non-overlapping in time. In fact, this is also true when task cues and stimuli are non-overlapping in space (i.e., where the stimuli are distanced from the cues rather than being placed more closely together). Thus, lag-2 repetition costs only occur when cues and stimuli are spatiotemporally integrated.

A second experiment demonstrated that this integration effect is driven by the lag-2 trial alone: temporal overlap between cues and stimuli on lag-1 (and even on the current trial) has no effect on lag-2 repetition cost. All that matters for lag-2 repetition costs is that cues and stimuli on the lag-2 trial were spatially or temporally integrated.

These findings don’t make sense from a backwards inhibition account of lag-2 repetition cost. Why would one incur a reaction time cost in switching back to a previously performed task only when that task and its cue had been presented close in time or space? In other words, shouldn’t subjects have to inhibit task A regardless of the relationship between cues and stimuli? Why should there be no lag-2 repetition cost when cues and stimuli are not integrated on those lag-2 trials?

The authors counter with what I’ll call the “turtles all the way down” hypothesis. They argue that backward inhibition is indeed still present, but is actually working double-time: on the lag-1 trial, they claim that subjects are inhibiting both possible tasks! Thus, according to this account, there is no difference in reaction time to complete A when comparing task orders A-B-A and C-B-A, since task A is inhibited regardless of the lag-2 trial.

This explanation is a little unsatisfactory for several reasons. First, there’s reason to believe that backwards inhibition itself is an unnecessary construct to explain lag-2 repetition cost. Therefore, explaining an absence of lag-2 repetition cost with not one but two backward inhibition processes simply proliferates an already-unnecessary construct.

Second, all of these results can also be interpreted within an alternative framework. The idea here is that to switch to task B, patterns of neuronal activity must change from a stable representation of the previous task to the representation of task B. This process involves attractor dynamics, such that the attractor basin for task B is widened in the direction of the previous task. When one must later switch back to this previous task, the representation of this previous task now partially belongs to the attractor basin for task B, and one encounters difficulty escaping this attractor basin (i.e., switching back to the lag-2 task). Thus, an absence of lag-2 repetition cost would occur when the task B attractor basin is less deformed. This could be expected to happen in several situations:

1) when the first task is completed relatively quickly, less of an attractor is established, meaning that the second task’s attractor basin is less affected.

2) when the first task cue is presented for a longer period of time, as in the current experiment, this creates a stronger attractor for the first task, which is then harder to escape to switch to task B. This results in a more deformed attractor on task B, yielding a lag-2 repetition cost.

Although this explanation probably sounds very abstract, it has several advantages. It can be implemented in standard self-recurrent network models, and is therefore consistent with many theories about task set representation in prefrontal cortex. It also highlights one shortcoming of the previous work: cue-stimulus integration was here confounded with cue display time. Finally, it makes some new predictions: lag-2 repetition cost should not be dependent on cue-stimulus integration when cue display time is equated between conditions; reaction times on current lag-2 repetition trials should be slower when the lag-2 trial had been completed more slowly; lag-1 reaction time should likewise be slower when lag-2 trials are completed more quickly.

So, perhaps if the phone rings after you’ve decided to take a break from that email, you’re just as fast to continue writing it after the call than if you’d been doing something else. However, the import of this fairly academic argument is not merely in optimizing productivity, but understanding how we control our behavior under situations of conflict. According to the backwards inhibition view, you are constantly inhibiting previous experiences in order to achieve your current goal. According to the attractor framework I’ve described, any slowing reflects a more passive interference process, where goal representations are subtly influenced by the efficiency of your other goal representations and the amount of time they are activated.