Non-Inhibitory Accounts of Negative Priming
The concept of "inhibition" is central to cognitive sciences, and yet some have raised doubts about the usefulness of this psychological construct. In their excellent 2003 chapter "In Opposition to Inhibition," MacLeod, Dodd, Sheard, Wilson and Bibi review the history of the construct (dating back at least to William James in 1890) and describe the experiments in which "inhibition" is used to explain the results. Below, I summarize their complaints about the use of this construct in the context of negative priming, as well as their proposed non-inhibitory accounts for this phenomenon.
Negative Priming
In the Stroop task, subjects must name the ink color of words like "RED;" subjects are much slower to name the ink color of GREEN if preceeded by RED than if preceeded by YELLOW. This has been interpreted as reflecting inhibition of the concept "red" so that the ink color can be named, which translates into slowed reaction time on a subsequent trial where the concept "red" must become activated. This slowed reaction time is called negative priming.
MacLeod et al review an alternative account of this effect which seems much more in line with current evidence. The "feature mismatch" account states that slowed reaction time reflects conflict arising from the stimulus itself rather than the response that is required. According to this view, it takes longer for the subject to process the stimulus if it shared overlapping yet mismatching features with the stimulus on the previous trial. So, in this case it takes longer to name the ink color of GREEN after naming the ink color of RED because the stimuli overlap in features - in terms of the concept "red" - but this feature overlap is mismatching in terms of how it relates to the first and second stimuli. Specifically, "red" refers to the ink color of the first stimulus, but the word name of the second stimulus.
Additional support for the idea that negative priming results from feature mismatch are several demonstrations that negative priming disappears altogether if the stimulus features do not mismatch between the previous trial and the current trial - regardless of whether the response requires attending to previously ignored information. To be perfectly clear, MacDonald & Joordens showed that "stimulus features" can include a visible feature of the stimulus itself or the features/semantic associates used to selectively attend to them.
For example, in a verbal task where subjects are presented with two words (MOUSE & AMERICA) and must pick the larger object, large negative priming (over 5 times the standard effect) is observed if the subsequent trial contains MOUSE & FLEA. In this case, the selection feature of "MOUSE" is [smaller] on the first trial and [larger] on the second trial, thereby creating a mismatch.
If the subsequent trial involved picking the smaller of WORLD & AMERICA, again there is a selection feature mismatch: the selection feature for AMERICA was [larger] on the first trial, and [smaller] on the second trial. Again, here you see large negative priming (which is impressive given that subjects usually show positive priming when making the same response twice in a row).
In contrast, no negative priming is observed if the first trial involves picking the larger of MOUSE & AMERICA and the subsequent trial involves picking the smaller of MOUSE & AMERICA. In this case, the selection feature of MOUSE is [smaller] both on the first and second trials, and thus there is no negative priming. In contrast, inhibition accounts would predict slowed reaction time on any trial where the subject is required to attend or respond to previously ignored information.
This demonstrates that mismatching selection features are the source of reaction time slowing, not difficulty in disinhibiting the previously ignored concept, or difficulty at the response level. Note that this effect has been replicated with a variety of types of stimuli mismatch, including physical size, numerical magnitude, and word color.
Conclusions
As argued by MacLeod et al. and as demonstrated by MacDonald & Joordens, negative priming can be explained as a result of stimulus feature mismatch between previous and current items. Whether those features were "ignored" to give a correct response seems immaterial; the important determinant in reaction time slowing is whether the features associated with a stimulus are congruent or incongruent with the features/associations activated on a previous trial.
It is interesting to note that while current computational models of the Stroop task do not explicitly model trial sequence effects of the kind that give rise to negative priming in Stroop (nor, for that matter, do they address increased errors/reaction time on incongruent trials following congruent trials, a phenomenon sometimes attributed to goal neglect), relatively minor modifications may allow them to successfully simulate these sequence effects.
Consider the case presented at the start of this post, where a neural network would process two trials in succession: GREEN and then RED. If activation is not "zeroed out" between events, lingering activity in the "red" output unit may cause slowed reaction time (aka cycles to settle) on the second trial, where the correct output is "blue." This would be a successful simulation of negative priming without recourse to directed inhibition, in line with the proposal of MacLeod et al. However, this might be argued to reflect mismatch at the output level, whereas Macdonald & Joordens showed that mismatch effects occur at the level of stimulus processing. It is not clear that current network models can account for this effect.
Consider another case, where the network would process GREEN and then RED. Here, the first trial could contribute to a "word reading" task unit to become more strongly active, which would linger into the second trial . This lingering activity could either result in either an outright error ("red") or would slow the number of cycles to settle on the correct output ("blue") because the "color naming" task unit would experience more competition. This would be a successful simulation of goal neglect.
Negative Priming
In the Stroop task, subjects must name the ink color of words like "RED;" subjects are much slower to name the ink color of GREEN if preceeded by RED than if preceeded by YELLOW. This has been interpreted as reflecting inhibition of the concept "red" so that the ink color can be named, which translates into slowed reaction time on a subsequent trial where the concept "red" must become activated. This slowed reaction time is called negative priming.
MacLeod et al review an alternative account of this effect which seems much more in line with current evidence. The "feature mismatch" account states that slowed reaction time reflects conflict arising from the stimulus itself rather than the response that is required. According to this view, it takes longer for the subject to process the stimulus if it shared overlapping yet mismatching features with the stimulus on the previous trial. So, in this case it takes longer to name the ink color of GREEN after naming the ink color of RED because the stimuli overlap in features - in terms of the concept "red" - but this feature overlap is mismatching in terms of how it relates to the first and second stimuli. Specifically, "red" refers to the ink color of the first stimulus, but the word name of the second stimulus.
Additional support for the idea that negative priming results from feature mismatch are several demonstrations that negative priming disappears altogether if the stimulus features do not mismatch between the previous trial and the current trial - regardless of whether the response requires attending to previously ignored information. To be perfectly clear, MacDonald & Joordens showed that "stimulus features" can include a visible feature of the stimulus itself or the features/semantic associates used to selectively attend to them.
For example, in a verbal task where subjects are presented with two words (MOUSE & AMERICA) and must pick the larger object, large negative priming (over 5 times the standard effect) is observed if the subsequent trial contains MOUSE & FLEA. In this case, the selection feature of "MOUSE" is [smaller] on the first trial and [larger] on the second trial, thereby creating a mismatch.
If the subsequent trial involved picking the smaller of WORLD & AMERICA, again there is a selection feature mismatch: the selection feature for AMERICA was [larger] on the first trial, and [smaller] on the second trial. Again, here you see large negative priming (which is impressive given that subjects usually show positive priming when making the same response twice in a row).
In contrast, no negative priming is observed if the first trial involves picking the larger of MOUSE & AMERICA and the subsequent trial involves picking the smaller of MOUSE & AMERICA. In this case, the selection feature of MOUSE is [smaller] both on the first and second trials, and thus there is no negative priming. In contrast, inhibition accounts would predict slowed reaction time on any trial where the subject is required to attend or respond to previously ignored information.
This demonstrates that mismatching selection features are the source of reaction time slowing, not difficulty in disinhibiting the previously ignored concept, or difficulty at the response level. Note that this effect has been replicated with a variety of types of stimuli mismatch, including physical size, numerical magnitude, and word color.
Conclusions
As argued by MacLeod et al. and as demonstrated by MacDonald & Joordens, negative priming can be explained as a result of stimulus feature mismatch between previous and current items. Whether those features were "ignored" to give a correct response seems immaterial; the important determinant in reaction time slowing is whether the features associated with a stimulus are congruent or incongruent with the features/associations activated on a previous trial.
It is interesting to note that while current computational models of the Stroop task do not explicitly model trial sequence effects of the kind that give rise to negative priming in Stroop (nor, for that matter, do they address increased errors/reaction time on incongruent trials following congruent trials, a phenomenon sometimes attributed to goal neglect), relatively minor modifications may allow them to successfully simulate these sequence effects.
Consider the case presented at the start of this post, where a neural network would process two trials in succession: GREEN and then RED. If activation is not "zeroed out" between events, lingering activity in the "red" output unit may cause slowed reaction time (aka cycles to settle) on the second trial, where the correct output is "blue." This would be a successful simulation of negative priming without recourse to directed inhibition, in line with the proposal of MacLeod et al. However, this might be argued to reflect mismatch at the output level, whereas Macdonald & Joordens showed that mismatch effects occur at the level of stimulus processing. It is not clear that current network models can account for this effect.
Consider another case, where the network would process GREEN and then RED. Here, the first trial could contribute to a "word reading" task unit to become more strongly active, which would linger into the second trial . This lingering activity could either result in either an outright error ("red") or would slow the number of cycles to settle on the correct output ("blue") because the "color naming" task unit would experience more competition. This would be a successful simulation of goal neglect.
posted by Chris Chatham at 8:39:00 AM. Now go to the current Developing Intelligence blog!
2 Comments:
I'm not in your field so this is a bit technical for me, and my thoughts may be off base.
But there seem to be two interesting ideas that might apply.
#1) Recent stimuli leave a degree of activiation, which in my layman's terms would help to provide the context in which we interpret current stimuli.
For example if we've just been talking about airports, I interpret "flight" differently than if we've just been talking about prisons.
#2) When the current state of activation is ambiguous, i.e has more than one interpretation, responses are slower, and confusion greater.
Hey, thanks for leaving these very interesting comments!
You are absolutely correct on both accounts. Interestingly, these characteristics of neural processing are often avoided by connectionist models, which tend to zero-out activation between (what they consider to be) discrete events.
In this sense, the "dynamical systems" modeling tradition is definitely superior, in that "nested timescales" are acknowledged as a reality that's worth modeling.
Search google for dynamical system + modeling and you'll likely come up with some interesting stuff.
Post a Comment
<< Home