Can you truly ignore anything? New evidence questions "filtering" accounts of memory

A number of previous behavioral and neuroimaging experiments, as well as computational models, support the idea that people can filter the contents of memory and perception so as to focus on only the information that's currently relevant. For example, in a visually-complex environment, distracting items often go unprocessed by the human visual system, perhaps due to some low-level perceptual bandwidth limitations. However, in cognitively-complex environments (when working memory, but not perception, is loaded to capacity), distractors are more likely to be processed (this is Lavie et al's work, discussed previously). Similar effects are seen in event-related potentials and fMRI, such that distractors are more likely to be processed by those of low-working memory capacity. This could suggest that mnemonic and perceptual filtering may be enabled by the same high-level cognitive processes involved in working memory.

However, a recent article (titled "The contents of visual memory are only partly under volitional control") by Olson, Sledge Moore, and Drowos casts some doubt on this interpretation. In a series of experiments, the authors demonstrate that distractors always enter memory so long as subjects are restricted from using spatial attention to focus on particular parts of the environment. Evidently, one need not invoke some complex high-level capacity for filtering perception to explain previous results, but merely posit that subjects were shifting their attention to different parts of the display in order to determine what to encode.

Olson, Sledge Moore & Drowos demonstrated this in six experiments in which a variable number of shapes were presented sequentially; subjects were asked to remember "target" shapes surrounded by a white box, and to indicate whether a subsequently-presented "probe" image matched any of the preceeding targets. On some trials, distractor shapes (not surrounded by a white box) were also presented. In every experiment, subjects were significantly less accurate in rejecting non-matching items when they had appeared as distractors, relative to when no distractors were present.

But what's really interesting here is the robustness of this phenomenon. Distractors reduced subjects' ability to reject non-match items:

  • regardless of memory load (i.e., whether subjects had been asked to remember 2 or 4 items)
  • regardless of whether the to-be-remembered items were abstract shapes or faces
  • regardless of whether subjects were warned in advance if an upcoming item would be a target

The only manipulation that affected the intrusion of distractors in memory was the simultaneous presentation of items, in which subjects could easily use spatial attention to attend only to targets. When the size and duration of this display was reduced to preclude spatial shifts in attention, distractors once again entered memory.

One interpretation of these results is that subjects used a long-term memory strategy to assist in the task. The novel shapes used in most of the experiments may not have had strong working memory representations, which are thought to build up over repeated experiences; instead, medial memory systems like the hippocampus are thought to be more important for the kind of rapid learning and memory for novel items required here. This is consistent with evidence that working memory capacity (measured with Cowan's K, as in some previous work using novel stimuli) was unaffected by the presence of distractors (except for when faces were used - a non-novel stimulus). It is also consistent with the finding that distractors were still well-recognized in a subsequent surprise test of long-term memory. Surprisingly, it is precisely this explanation that the authors caution against, pointing to a long-standing debate about how cleanly these two memory systems can actually be dissociated. I have trouble seeing how a unitary model of memory can account for the absence of an effect of distractors on Cowan's K, but acknowledge that distinctions between working and long-term memory can be murky.

I find these results incredibly exciting, because they conflict with a number of previous studies of similar phenomena. Vogel and colleagues have shown that distractors do not enter working memory when the items are easily recognized (colored, oriented bars), even if shifts in spatial attention are precluded (by displaying items for only 100ms, the same display duration used in one of Olson et al's conditions). My guess is that the novelty and complexity of the to-be-remembered items is the key to this particular discrepancy. Colored, oriented bars are more likely to have a robust WM representation and much less likely to rely so strongly on the hippocampal or long-term memory representations likely taxed in the current study, and thus the former are much more likely to be successfully filtered via top-down attentional set. If anything, this explanation of the discrepancy actually supports a dual-process view of memory.

An alternative explanation of these results is that subjects failed to "bind" the visual stimuli with their target status (indicated by a white box surrounding the item) and thus were unable to filter the contents of memory on the basis of which were distractors and which were targets. In support of this explanation, previous work has shown that sequential stimulus presentation does not allow subjects to reliably associate stimulus features belonging to the same object (i.e., visual "binding").

Another alternative explanation is that in the previous work demonstrating perceptual filtering (e.g., the Vogel et al paradigms), subjects were spatially shifting their attention on a iconic memory trace of the preceding stimuli. Displays in those paradigms are not visually masked, leading to the possibility that spatial attention accomplishes filtering in those circumstances. The alternative possibility, more in line with computational models of "filtering," is that perceptual features of the stimuli are used directly by the basal ganglia in determining the likelihood that they will be updated into working memory.

To summarize, Olson, Sledge Moore & Drowos have demonstrated that distractors intrude into "memory" unless spatial attention precludes their encoding. This intrusion is unaffected by top-down attentional set, at least when novel stimuli are used; on the other hand, this intrusion does not appear to affected indices of working memory (Cowan's K). One possibility is that the distractors enter long term memory, as reflected by performance on a surprise test of recognition. More speculatively, filtering of distractors from memory may be limited to cases in which distractor status can be "visually bound" to each item, or limited to cases in which objects are less novel, or both.

More like this

From the abstract:
When attention coupled with eye movements could be used to select targets, distractors were no longer encoded into memory. When eye movements were constrained, distractors once again intruded into memory. These findings suggest that top-down control processes are insufficient to filter the contents of visual memory.

But eye movements are top-down control processes that filter the contents of visual memory.

(Well, except that "filtering" is a really misleading metaphor that has been screwing up the psychology of attention for decades - "selecting" would be way better. But really! You stop people using one of their most important attentional mechanisms and attention no longer does its job properly? What a surprise!)

When we look around within a visual scene, is visual information automatically placed in visual memory during each saccade, or can we control which information is retained and which is excluded?

Yes it is, and yes we can.® It is not one or the other. We control our saccades, and they are one of the most important mechanisms through which the contents of visual memory are regulated.

(I do not doubt that there are other, internal, attentional mechanisms that make further selections from the information that gets past the eyes and into the brain, but the talk of retaining and excluding is messed up. Some of the available information gets used because it represents what we need to know to accomplish whatever our current project may be, and the rest does not get used, because it represents irrelevant stuff. There is no need to deliberately exclude stuff, you, or rather your cognitive system, just doesn't choose to do anything with some of the data that is available. But of course, if you cripple the selection mechanisms through some experimental manipulation, the wrong stuff gets selected, and whatever purpose it was being selected for - maybe some experimentally imposed task - gets derailed.)