Chapter 2 Bottlenecks, resources, and capacity
Quickly, what is fourteen times thirteen? Calculating that in your head takes a while, at least a few seconds. And if I set you two such problems rather than just one, I’m confident that you would do those problems one at a time. Our minds seem to be completely incapable of doing two such problems simultaneously (Oberauer 2002; Zylberberg et al. 2010). This limitation is remarkable given that each of our brains contains more than 80 billion neurons. The problem is not a lack of neurons, really, but how they are arranged - our mental architecture.
Multiplying and dividing two-digit numbers may not be something you attempt to do every day. You might think, then, that if you were doing lots of such problems each day, you could eventually do more than one at a time. This is probably wrong - consider that a task we do have daily practice with is reading. Despite years of reading dozens if not hundreds of words a day, the evidence suggests that humans can read at most only a few words at a time, and some research further indicates that we can really only read one word at a time (Alex L. White, Palmer, and Boynton 2018; Reichle et al. 2009). At least some of the bottlenecks of human information processing, then, appear to be a fixed property of our processing architecture.
To flesh out what I mean by ‘bottleneck’ here, consider a standard soft drink bottle. If you invert a full bottle, most of the liquid volume will be pressing down on the neck. The narrowness of the neck restricts the rate at which the liquid can exit the bottle. Similarly, a large volume of signals from sensory cortex ascending the cortical hierarchy press up against higher areas that are more limited in capacity.
The parallel processing happening in visual cortices, such as the multiple neurons dedicated to each patch of the visual field, gets a number of tasks done, so that higher stages don’t have to do those tasks. These tasks appear to include the encoding of motion direction, color, and orientation throughout the visual field. Local and regional differencing operations happen for those features, resulting in salience, whereby odd features become conspicuous in our visual awareness. In the below display, for example, you should be able to find the blue objects very quickly.
Figure 4: Thanks to featural attention (to color in this case), you should be able to find the blue circles very quickly.
For other judgments, higher, post-bottleneck brain areas that are very limited in capacity are critical. The visual word form area in the occipitotemporal sulcus of the left hemisphere, that seems to be needed to recognize words, is one example (Alex L. White et al. 2019). Being limited in processing capacity to just one stimulus, the word recognition will not happen in a crowded scene until something selectively directs the visual signals from a word to the visual word form area. We often use the term selective attention to refer to this “something” that directs particular visual signals to the bottlenecks of limited-capacity processes. If there were no bottlenecks, there would be no need for selection for cognition (selection would be required when an action needed to be chosen).
So far the picture I have painted has been one of a torrent of visual signals impinging on a narrow bottleneck of signals that continue onward. But cortical processing is rarely a one-way street, and the way visual attention works is no exception. Visual attention seems to work partly by biasing processing within visual cortices, rather than leaving that unchanged and blocking all but a few signals at a later bottleneck stage. Thus, processing capacity may be restricted by limitations on control signals from high-level (possibly parietal) cortex that restrict processing capacity, as well as the more familiar idea of a structural bottleneck where ascending visual signals reach a lower-capacity neural mechanism.
To the extent control signals are a limitation, a resource metaphor can be apt. The control of selection may reflect a finite pool of neural resources in parietal cortex that bias which visual signals are cognitively processed. Thus I will sometimes use the term “limited resource” when referring to how we are restricted in how many visual representations are processed.
The word “resource” carries the appropriate connotation that people can choose how to apply their finite processing capacity; ordinarily a resource is something that can be used in different ways. For example, the term suggests that one might use three-quarters of one’s processing capacity for one target while using the other quarter for a second target. And indeed, there is evidence that people can favor one target over another when tracking both (Chen, Howe, and Holcombe 2013; Crowe et al. 2019).
While word recognition seems to be able to process only one stimulus at a time, other visual judgments may be limited in capacity relative to massively parallel sensory processing, but have a capacity greater than one. Object tracking seems to be one such ability. People appear able to track more than one target at the same time, although researchers haven’t fully ruled out the possibility that tracking multiple objects happens via a one-by-one process that rapidly switches among the tracked objects.
The existence of processes with a capacity of just one object (I will introduce the term “System B” for this in Section 6) is a good reason to have a process that can keep track of the location of important objects in a scene. We are then always ready to rapidly shunt a subset of them to higher-level processing, rather than having to search for it.