2/02/2006

The Mind's Eye: Models of the Attentional Blink

In the attentional blink paradigm, two targets are displayed in rapid serial visual presentation; astonishingly, participants show a brief temporal window in which they cannot identify the second target, while on either side of that window recognition proceeds normally. It is as though the proverbial mind's eye must "blink" in order to attend to two temporally distinct meaningful items. Imaging research shows that the second incoming item is still processed by higher visual areas, even if subjects are unable to report seeing that item; so where is the neural activity that determines whether we are aware of an object? In the pursuit of this exact question, many researchers interested in consciousness have developed computational models of the attentional blink.

One mechanism that may give rise to the attentional blink is that both targets (T1 and T2) can be processed in parallel if presented close enough together; otherwise serial processing takes over, which creates a temporary bottleneck for subsequent processing of T2 from visual short-term memory. An alternative account posits that the bottleneck occurs in a limited-capacity processing stage required for awareness and preparation of a response. There are also hybrid models that incorporate both of these ideas.

A second mechanism that may underlie attentional blink is the backward masking of visual stimuli, such that each successive image interferes with the processing of the previous image at a perceptual level. This is reflected by the fact that when the images in the serial presentation are separated by blank pages, the timeline of the attentional blink is shortened.

In their computational modeling of the attentional blink, Bowman and Wyble invoke a distinction between type and token such that types record the specific semantic and perceptual features of an item, while the token information represents the information specific to each presentation (for example, duration and temporal position). Their model uses a fairly standard sigmoid activation function (incorporating terms for unit bias, excitation, inhibition and leak currents) and a structure that involves two stages. In the first stage, localist representations of incoming stimuli are subject to masking through lateral and feedforward inhibition in two layers, which then project to two more layers with slow fading functions. The first is intended to represent the items at a semantic level, and the second is intended to represent the items at a categorical level, to which there is input from a "task demand" unit. This layer is known as the "task selection layer," consisting of recurrent excitatory connections within units and weak inhibition among units, resulting in interference between simultaneously active representations. Further, a 50ms pulse of excitatory activity throughout these two layers occurs whenever a unit in the layer is activated, meant to simulate the brief salience that an attentional window is assumed to bestow on representations it selects. The researchers also posit a "shutoff" layer which can inhibit each neuron in the task selection layer after a "sufficient" amount of activity; this is critical for both ensuring that each token is encoded by only one type, but also for creating repitition blindness. Each token has four states, and at the end of the serial presentation, the model performs a recall phase to disambiguate the order of presentation (but this mechanism is not specified in the model), which incorporates a random number between 0 and 1 to calculate the probability of an order inversion on that trial.

In this model, attentional blink emerges as a result of the suppression of recurrent excitation in the task-selection layer from both lateral inhibition and the inhibition that occurs during the tokenization process; when the temporal conditions are right, T1 binding takes long enough that the subsequent T2 representation has faded before it could be activated sufficiently for tokenization. Lag-1 sparing occurs when T1 and T2 are encoded into the same token, using the same excitatory pulse; and backward unmasking effects occur because T1 can then be bound more rapidly.

In a competing computational theory of the attentional blink, Nieuwenhuis, Gilzenrat, Homes & Cohen propose that refractoriness in the locus coeruleus (LC) norepinephrine system may be to blame for the attentional blink. The LC is a brainstem region containing half of all noradrenergic neurons in the CNS which has been implicated in modulating arousal level, affective state, and the sleep-wake cycle in primates. LC also shows phasic activity increases during attentional processing, with a subsequent refractory period that closely corresponds with the timeline of attentional blink. Interestingly, the LC shows increased tonic levels of activity when primates are disengaged from a task, and increased phasic (50-100ms) levels only when the animal is activately attending; this modal shift in firing patterns is thought to increase the signal-to-noise ratio of subsequent firing. Release of nor-epinephrine (NE) by LC is excitatory in cortex but actually autoinhibitory on the LC for about 400-450ms poststimulus; this is the mechanism causing both the refractoriness and the attentional blink.

Accordingly, MEG and ERP data show that magnitude of LC activity correlates with the size of the attentional blink, and empirical results suggest that the refractory period also correlates with the strength and duration of each stimulus (which may explain backward masking). This region has been called "the blaster" because it conveys salience to processing of task-relevant items, perhaps similar to the 50ms excitatory pulse in Bowman and Wyble's model. Neither of these models explicitly account for increased attentional blink following emotional stimuli, nor the lack of lag 1 sparing seen in many studies where the spatial location of images changes with order of presentation.

Both models differ significantly from Dehaene & Sergent's 2003 model in which targets become accessible through gamma-band oscillations between higher and sensory areas of cortex (although it's possible that the "blasters" in both models could trigger oscillatory activity in this frequency range). While the Bowman and Wyble model accounts for temporal inversion (the Niewenheiss et al model completely ignores this phenomenon), Bowman and Wyble do so at the expense of parsimony: it's not clear how biologically plausible their token/type distinction nor the shutoff layer may be, both of which are self-admittedly the heart of their model. In other words, the model has "done the job," but it's not at all clear that the brain does it in the same way.

Related Posts:
Anticipation and Synchronization (possible role of gamma-band oscillations)
Active Maintenance and The Visual Refresh Rate (visual refresh as distinct from 'cycle speed')
Kosslyn's Cognitive Architecture (of vision)
Selection Efficiency and Inhibition (is information inhibited by, or selected by working memory)

If you liked this, don't forget to digg it.

References:

Bowman H and Wyble B. Computational modelling of the attentional blink. In Angelo Cangelosi, Guido Bugmann, and Roman Borisyuk, editors, Proceedings of the Neural Computation and Psychology Workshop, volume 9, January 2005.

Dehaene, S., Sergent, C., & Changeux, J.-P. (2003). A neuronal network model linking subjective reports and objective physiological data during conscious perception. Proceedings of the National Academy of Sciences,
USA, 100, 8520–8525


Marois, R., Yi. D.-J., & Chun, M. M. (2004). The neural fate of consciously perceived and missed events in the attentional blink. Neuron, 41, 465-472.

Nieuwenhuis, S., Gilzenrat, M.S., Holmes, B.D., & Cohen, J.D. (2005). The role of the locuscoeruleus in mediating the attentional blink: A neurocomputational theory. Journal ofExperimental Psychology: General, 134, 291-307

Sergent, C, Baillet, S., & Dehaene, S. Timing of the brain events underlying access to consciousness during the attentional blink.. Nat Neurosci, September 2005

4 Comments:

Blogger Chris Chatham said...

Thanks to Tim for pointing me towards Bowman and Wyble's very interesting model!

2/02/2006 08:06:00 AM  
Anonymous Anonymous said...

Yeah, the model proposed by Bowman & Wyble is interesting, but I think its flawed in a deep way. In particular, the explanation of lag 1 sparing as being the result of encoding T1 and T2 into the same token is not at all in line with the type-token distinction outlined by Chun & Potter. Once something has formed one token, it cannot be seperated again -- thats the entire definition of a token!

I've raised this point with Brad Wyble before, but never in a context where we really got to chat about it. Either way, all the models of AB are interesting in their own way. I prefer to think of it as a parallel/serial thing, with a conveyor belt (I believe there is a picture of a model like this in Chun & Wolfe's 'Visual Attention' chapter from 2001; probably elsewhere as well), which gets you both AB and RB (repetition blindness).

2/02/2006 09:08:00 AM  
Blogger Chris Chatham said...

Yeah, I had some trouble with the "type/token" thing ... and this is apparently the more biologically constrained of their AB models.

AB and similar phenomena are great opportunities for computational modelers: whereas multiple system-level descriptions of the brain might be able to give rise to similar functionality, how many can lay claim to glitching in the same way the brain does?

I guess this is the same logic behind "simulated lesioning"...

2/02/2006 02:42:00 PM  
Blogger Brad Wyble said...

Tim, you're correct that our model's implementation, which combines two types into one token is questionable. We were willing to think this might be the case for just 2 targets, but recent data (Olivers Van der Stigchel & Hulleman 2007) has shown that sparing can be spread across 3-4 consecutive targets, forcing us to reevaluate the implementation of tokens.

I've just about finished an extension of the model which addresses your concern; now there is always 1 type per token.

This model replicates the new spreading of sparing data (also cueing, and whole report). It's basic character is similar to my original STST implementation, but the token binding system has been slightly modified.

Hopefully it will be in-press in a few months, and I'll present it at VSS this year (2007). Until then, if you have any questions, drop me a line at bwyble @ gmail[dot]com.

-Brad Wyble

3/31/2007 08:16:00 AM  

Post a Comment

<< Home