Shared structure facilitates working memory of multiple sequences via neural replay

Reviewer #1 (Public Review):

Summary:
Huang and Luo investigated whether regularities between stimulus features can be exploited to facilitate the encoding of each set of stimuli in visual working memory, improving performance. They recorded both behavioural and neural (EEG) data from human participants during a sequential delayed response task involving three items with two properties: location and colour. In the key condition ('aligned trajectory'), the distance between locations of successively presented stimuli was identical to their 'distance' in colour space, permitting a compression strategy of encoding only the location and colour of the first stimulus and the relative distance of the second and third stimulus (as opposed to remembering 3 locations and 3 colours, this would only require remembering 1 location, 1 colour, and 2 distances). Participants recalled the location and colour of each item after a delay.

Consistent with the compression account, participants' location and colour recall errors were correlated and were overall lower compared to a non-compressible condition ('misaligned trajectory'). Multivariate analysis of the neural data permitted decoding of the locations and colours during encoding. Crucially, the relative distance could also be decoded - a necessary ingredient for the compression strategy.

Strengths:
The main strength of this study is a novel experimental design that elegantly demonstrates how we exploit stimulus structure to overcome working memory capacity limits. The behavioural results are robust and support the main hypothesis of compressed encoding across a number of analyses. The simple and well-controlled design is suited to neuroimaging studies and paves the way for investigating the neural basis of how environmental structure is detected and represented in memory. Prior studies on this topic have primarily studied behaviour only (e.g., Brady & Tenenbaum, 2013).

Weaknesses:
The main weakness of the study is that the EEG results do not make a clear case for compression or demonstrate its neural basis. If the main aim of this strategy is to improve memory maintenance, it seems that it should be employed during the encoding phase. From then on, the neural representation in memory should be in the compressed format. The only positive evidence for this occurs in the late encoding phase (the re-activation of decoding of the distance between items 1 and 2, Fig. 5A), but the link to behaviour seems fairly weak (p=0.068). Stronger evidence would be showing decoding of the compressed code during memory maintenance or recall, but this is not presented. On the contrary, during location recall (after the majority of memory maintenance is already over), colour decoding re-emerges, but in the un-compressed item-by-item code (Fig. 4B). The authors suggest that compression is consolidated at this point, but its utility at this late stage is not obvious.

Impact:
This important study elegantly demonstrates that the use of shared structure can improve capacity-limited visual working memory. The paradigm and approach explicitly link this field to recent findings on the role of replay in structure learning and will therefore be of interest to neuroscientists studying both topics.

Reviewer #2 (Public Review):

Summary:
In this study, the authors wanted to test if using a shared relational structure by a sequence of colors in locations can be leveraged to reorganize and compress information.

Strength:
They applied machine learning to EEG data to decode the neural mechanism of reinstatement of visual stimuli at recall. They were able to show that when the location of colors is congruent with the semantically expected location (for example, green is closer to blue-green than purple) the related color information is reinstated at the probed location. This reinstatement was not present when the location and color were not semantically congruent (meaning that x displacement in color ring location did not displace colors in the color space to the same extent) and semantic knowledge of color relationship could not be used for reducing the working memory load or to benefit encoding and retrieval in short term memory.

Weakness:
The experiment and results did not address any reorganization of information or neural mechanism of working memory (that would be during the gap between encoding and retrieval). There was also a lack of evidence to rule out that the current observation can be addressed by schematic abstraction instead of the utilization of a cognitive map.
The likely impact of the initial submission of the study would be in the utility of the methods that would be helpful for studying a sequence of stimuli at recall. The paper was discussed in a narrow and focused context, referring to limited studies on cognitive maps and replay. The bigger picture and long history of studying encoding and retrieval of schema-congruent and schema-incongruent events is not discussed.