Really neat paper showing that memories reactivated during sleep with TMR can be decoded during sleep. As the intro highlights, this has been a major goal within sleep and memory research recently, and has been shown a few times very recently. Yet all these instances of neural memory reactivation evidence have their own specifics methods and caveats, so nice to see another “version” of it. Another highlight of the paper is a replication of the relationship between sleep spindles and memory reactivation – something else this research group has been pushing as of late. While the latter is interesting, my personal interest (and focus of this post) is more about the ability to decode reactivated memories. I’m interested in this from a methods perspective for future research designs. (As a related aside, Jarrod gets referenced in the last line of paper – the non-monotonic stuff – as something that can be investigated with a neural index of memory reactivation during sleep.)
With the intro serving as a summary of all the recent TMR reactivation decoding work, I think it nicely highlighted the subtle variations across attempts. Obviously some approaches used EEG while others fMRI, but also differences like item- vs category- level decoding and whether the decoder was built/trained with waking or sleep data (of course always tested on sleep) are highly relevant. I think the latter is particularly relevant for discussion here, as in this paper they could decode left/right within sleep, but not when training on wake and testing on sleep. That being said, their decoder was trained on sleep data using features derived from waking data, so the two were not totally separated during analysis. I found the whole decoder rather strange, but perhaps it’s just unfamiliarity. They build an across-subject decoder, which in my experience is very rare (for reasons of difficulty due to inconsistent neural signals between subjects). This might be less of an issue with EEG – again I’m not so familiar – and perhaps the robust signals of motor lateralization make this easier, although in my brief delve into EEG left/right imagery decoding, I still came across many comments about the issue of across-subject decoding. In any case, I can only speculate that they did this due to the low number of trial counts that would have resulted for any within-subjects decoders (since they end up breaking trials down into strict accuracy-dependent categories). I wonder if this was planned, because I would imagine the initial plan was to use the waking data for training and sleep data for testing, in which case there would have been enough trials for each subject?? On that note, perhaps this across-subjects decoding is more common in TMR reactivation for this reason, that even with a wake-trained decoder, the TMR cue is only presented during sleep for a specific item a handful of times, so they need more test trials??
With the within-sleep across-subject decoder, they found L/R classifier discriminability roughly 1-4 seconds after the TMR cues. I think the highlight of the paper is the variety of relationships they show between neural data and memory performance. The classifier discriminability (i.e., L/R “decodability”) increased when they only included trials of memory items that were remembered before the nap, and increased from there when including only the subset of item that were retained from before to after the nap. Thus, decodability was related to memory performance.
As for the spindle stuff, it jives very well with recent work from both these labs. They find that spindles after the cue are good for memory and spindles before the cue are bad (presumably from post-spindle refractory period). While these effects seem a bit weak, I feel no concern over that at all given all the recent work from these labs.