thank you Andrey! agreed - TAFKAP is a great tool.

if you're referring to the two stimuli that subjects memorized on each trial, we retrocued one and used TAFKAP to reconstruct the motion direction of the retrocued item in the post-cue delay.

(3) there's more to a memory than there seems. using the mean of a distribution or averaging data over many trials obscures meaningful information about what the subject is maintaining.

11/11

(2) visual and parietal cortex store surprisingly rich information that is related to subjects' behavior. maybe working memory representations in these regions is not some fragile form of memory or epiphenomenal?

10/n

what does this tell us? (1) well, individual memories are rich - with memory and uncertainty jointly encoded. this can help us form the basis of theories for how working memory is converted into behavior - an area that we know very little about.

9/n

decoded neural distributions from visual and parietal cortex were also asymmetric! most importantly, the asymmetry of the neural distributions matched that of the behavioral distributions! this is a mutual validation of these single-trial approaches.

8/n

what's more, this asymmetry was meaningful - there was more information in their full probability distribution than you'd get with a single report. this means there's more to a working memory than we know if we only ask subjects to report a single value!

7/n

rather than the roughly normal distributions (von mises, if we're being technical) we're used to seeing for behavioral error, the probability distributions participants constructed on individual trials were highly asymmetric.

6/n

how did we get around this? we used a task that allowed subjects to construct a probability distribution reflecting their memory on teach trial. we paired this with single trial reconstructions of individual memories using fMRI and a bayesian decoder.

5/n

distinguishing between these alternatives is hard! if you aggregate data to get trial-averaged behavior or decoded memories, as most studies do, these two types of representations make identical predictions. you need rich, single-trial estimates of individual's memories.

4/n

this is an important distinction! information-rich representations jointly encode memory and its associated uncertainty, while information-sparse representations either do not account for uncertainty or encode it separately from the memory itself.

3/n

in a study led by ying zhou, we asked whether working memory representations are "information-rich" - full probability distributions over the space of possible values, or "information-sparse" - point estimates in feature space plus some notion of confidence.

2/n