Slots versus resources
3.3. The maintenance capacity for feature associations
In the preceding section on the maintenance of single features, we have broadly discussed the most actual hypothesis concerning the working memory capacity limit along with some points of discussion. The fractionation of the section on the maintenance of single and the maintenance of integrated features was rather chosen on structural grounds,
complying as such with the structure of the preceding chapters, than on theoretical grounds.
The capacity limit and accompanying issues discussed in the previous chapter might apply to a large extent as well to the maintenance of integrated features. We have described that for the multi-component and the TBRS model, the capacity limit of four items applies to the episodic buffer and for the embedded process model this limit concerns the focus of attention. These concern in both cases the memory structure accommodating integrated features. The capacity limit is hence acknowledged to apply to “items” rather than to features.
The origin of the capacity limit of about four items was largely based on experimental evidence from studies with single features, though initially this limit seemed to conform to the maintenance of objects as well (e.g., Luck & Vogel, 1997). This had already been suggested by Miller in 1956. He claimed the number of chunks of information one can maintain in immediate memory to be fixed, independently of the number of features a chunk contains.
Nevertheless, the capacity limit of four items for the maintenance of integrated features has received less unanimity than its capacity limit for the maintenance of single features. The next two sections will discuss a number of studies favoring or disfavoring this equivalent capacity limit for single and integrated features. We will follow the same structure as in chapter two and start by discussing research on the capacity limits for the maintenance of integrated features belonging to a same domain (i.e., the visuo-spatial or the verbal domain). Next, we will discuss the capacity limits for cross-domain features associations.
One of the most exemplar studies confirming the capacity limits for the maintenance of integrated features to be the same as the capacity limit for the maintenance of single features is the study by Luck and Vogel (1997), which we have detailled in chapter two.This
study showed that as much colored oriented bars could be maintained as single colors or single orientations. Vogel, Woodman, and Luck (2001) also showed that participants could maintain eight colors when these were presented as four objects, as compared to only four colors when these were presented in isolation. Capacity limits should thus be regarded in terms of objects according to these authors. As stated before, these studies have however severely been questioned. Several attempts to replicate the experiment with integrated color objects have failed (e.g., Delvenne & Bruyer, 2004; Olson & Jiang, 2002; Wheeler &
Treisman, 2002) and the fact that the associations between features had never been tested added doubt to the “object”-based maintenance of these feature associations.
More recently, several studies have reported evidence to the alternative hypothesis that fewer objects can be maintained than single features. For example, Allen et al. (2006)
compared color and shape feature recognition under different conditions, making use of a change detection paradigm. They presented four colors, four shapes or four colored shapes. In the case of the colored shape objects, participants had to maintain both the color and the shape as they did not know in advance on what feature dimension they would be tested. Though the corrected recognition score for colors was the same in the feature as in the object presentation condition, this score was better for shapes when only the shape had to be maintained. This design was thus similar to the one used by Luck and Vogel (1997), but the results were less convincing. A similar study by Cowan et al. (2013) even showed that the immediate memory capacity as measured by Cowan’s formula k was smaller for shape as well as for color
information when both had to be attended (M = 2.31) than when only attending to one feature (M = 2.91). The suggestion that as much objects as single features could be maintained was further discredited by studies showing that working memory maintenance performance declines when more features per object have to be maintained. Oberauer and Eichenberger (2013), for example, showed that change detection was worse for objects composed of six features instead of objects composed of three features, which were both worse than change detection for one-feature objects. Hardman and Cowan (2015) came to this same conclusion that increasing the number of features for an object reduces change detection. Additionally, it was shown in several studies that even though a certain number of features of an object might be memorized, the associations between these features are generally less well remembered than its constituent features (Cowan et al., 2013; Johnson et al., 2008; Morey & Bieler, 2013;
Vergauwe, Langerock, et al., 2014).
Nonetheless, should the idea of a same capacity limit for the maintenance of single features and objects be completely abandoned? Not per se. An intermediate account has been suggested by Cowan et al. (2013). These authors proposed the existence in working memory of three or four slots. These slots could be filled with either single features or objects. In theory, when presenting three, four, or even more objects, participants should be able to maintain three or four objects within working memory. Studies cited above (Allen et al., 2006; Hardman & Cowan, 2015; Oberauer & Eichenberger, 2013) had shown that this is not the case and this is also not what Cowan et al. (2013) had observed. They observed instead that on some occasions only the color of the object was maintained, or only the shape, while on other occasions both color and shape were maintained. When summing the number of objects for which at least one feature was maintained, this added up to about three or four items. So while objects may lose some of their features, the theoretical capacity limit remains the same for single features and objects. In order to validate this account, it should however be elucidated what conditions result in the loss of the feature information and which
information is most prone to this kind of loss. Nevertheless, this intermediate account offers hence a reasonable explanation in terms of a fixed capacity.
All studies discussed so far concerned feature associations within the visuo-spatial domain. Research on the capacity limits of verbal information has not elaborated on the maintenance of associations to a same extent. Studies regarding this capacity limit have mainly been executed by Cowan and his research team. In a typical experiment, Cowan, Chen, and Rouder (2004) asked participants to study a large number of associations between word pairs before the start of the actual experiment in order to turn these pairs into
predisposed chunks. In the actual experiment, participants were then presented with lists of four word pairs they had to remember. Recall of these lists was scored in terms of chunks.
One point was given if both words of the chunk were recalled, but as well if only one word was recalled. This scoring method allowed to observe a fixed capacity limit in terms of verbal chunks, which was comprised between three and four items as well. However, further studies identified delimitating conditions needed to observe this fixed capacity limit of three (rather than four) items. Chen and Cowan (2009) for example were able to confirm this capacity limit, but only when articulatory suppression was applied and order not taken into account.
Two more conditions were to be implemented in the study by Cowan et al. (2012) to observe the fixed capacity limit. First of all, there was the possibility for chunks to decompose. This latter condition implies that for example a two-word chunk could lose the connection between
the words and hence occupy two slots instead of one. Secondly, a contribution from long term memory had to be allowed.
Due to the involvement of long term memory, which influences additionally the chunking process, but also due to implicit linguistic restrictions on which words can follow each other, it is difficult to make adequate capacity estimates for verbal associations.
Nevertheless, Cowan and his team showed several results suggesting that the fixed capacity limit of three to four items might as well apply to verbal associations. Additional research based on different methodologies should however further confirm this result.
Research on the maintenance of within-domain associations has mainly compared the capacity limits between single features and feature associations by keeping constant the number of objects, while varying the number of features to be attended. The few studies on the capacity limits of cross-domain feature associations have instead rather focused on the number of features one could maintain when these features are presented isolated or
integrated. This was achieved by keeping the number of features constant, but manipulating their presentation mode. For example, the maintenance of four letters and four locations when presented as eight isolated features could be compared with the maintenance of four letter-in-location objects. This had been done by Prabhakaran et al. (2000) and Morey (2011). Both studies reported better recognition of integrated than separated features, though the
differences were rather small. However, this concerned only incidental binding of the
information in the integrated presentation condition. Morey (2011) broadened Prabhakaran et al.’s research design to include a condition obliging participants to maintain these bindings. A comparison between the amount of features maintained in this obligatory bound presentation and the separate presentation resulted in a much more pronounced difference, with feature maintenance much better in the case of the obligatory bound maintenance. Thus, these studies from Prabhakaran et al. (2000) and Morey (2011) show that feature maintenance is better when it is maintained as an object than as separated features. Unfortunately, pure capacity estimates in terms of the number of objects that could be maintained have to our knowledge not been reported yet for cross-domain feature associations.
The working memory capacity limits for the maintenance of single features are agreed on by a number of working memory theories and fixed at about three to four items.
Nevertheless, a number of specific conditions have to be met in order to actually observe this capacity limit. A fixation of the capacity limits for the maintenance of feature associations is less unanimously agreed upon. While several studies on feature associations comply with the capacity limit of three to four items, others have clearly shown this capacity limit to be lower.
Furthermore, capacity limits for the maintenance of cross-domain associations are clearly missing. Additional research is hence recommended.