From recency to central tendency biases in working memory: a unifying network model

Reviewer #1 (Public Review):

This paper aims to explain recent experimental results that showed deactivating the PPC in rats reduced both the contraction bias and the recent history bias during working memory tasks. The authors propose a two-component attractor model, with a slow PPC area and a faster WM area (perhaps mPFC, but unspecified). Crucially, the PPC memory has slow adaptation that causes it to eventually decay and then suddenly jump to the value of the last stimulus. These discrete jumps lead to an effective sampling of the distribution of stimuli, as opposed to a gradual drift towards the mean that was proposed by other models. Because these jumps are single-trial events, and behavior on single events is binary, various statistical measures are proposed to support this model. To facilitate this comparison, the authors derive a simple probabilistic model that is consistent with both the mechanistic model and behavioral data from humans and rats. The authors show data consistent with model predictions: longer interstimulus intervals (ISIs) increase biases due to a longer effect over the WM, while longer intertrial intervals (ITIs) reduce biases. Finally, they perform new experiments using skewed or bimodal stimulus distributions, in which the new model better fits the data compared to Bayesian models.

The mechanistic proposed model is simple and elegant, and it captures both biases that were previously observed in behavior, and how these are affected by the ISI and ITI (as explained above). Their findings help rethink whether our understanding of contraction bias is correct.

On the other hand, the main proposal - discrete jumps in PPC - is only indirectly verified.

The model predicts a systematic change in bias with inter-trial-interval. Unless I missed it, this is not shown in the experimental data. Perhaps the self-paced nature of the experiments allows to test this?

The data in some of the figures in the paper are hard to read. For instance, Figure 3B might be easier to understand if only the first 20 trials or so are shown with larger spacing. Likewise, Figure 5C contains many overlapping curves that are hard to make out.

There is a gap between the values of tau_PPC and tau_WM. First - is this consistent with reports of slower timescales in PFC compared to other areas? Second - is it important for the model, or is it mostly the adaptation timescale in PPC that matters?
Regarding the relation to other models, the model by Hachen et al (Ref 43) also has two interacting memory systems. It could be useful to better state the connection, if it exists.

Reviewer #2 (Public Review):

Working memory is not error free. Behavioral reports of items held in working memory display several types of bias, including contraction bias and serial dependence. Recent work from Akrami and colleagues demonstrates that inactivating rodent PPC reduces both forms of bias, raising the possibility of a common cause.

In the present study, Boboeva, Pezzotta, Clopath, and Akrami introduce circuit and descriptive variants of a model in which the contents of working memory can be replaced by previously remembered items. This volatility manifests as contraction bias and serial dependence in simulated behavior, parsimoniously explaining both sources of bias. The authors validate their model by showing that it can recapitulate previously published and novel behavioral results in rodents and neurotypical and atypical humans.

Both the modeling and the experimental work is rigorous, providing compelling evidence that a model of working memory in which reports sometimes sample past experience can produce both contraction bias and serial dependence, and that this model is consistent with behavioral observations across rodents and humans in the parametric working memory (PWM) task.

Evidence for the model advanced by the authors, however, remains incomplete. The model makes several bold predictions about behavior and neural activity, untested here, that either conflict with previous findings or have yet to be reported but are necessary to appropriately constrain the model.

First, in the most general (descriptive) formulation of the Boboeva et al. model, on a fraction of trials items in working memory are replaced by items observed on previous trials. In delayed estimation paradigms, which allow a more direct behavioral readout of memory items on a trial-by-trial basis than the PWM task considered here, reports should therefore be locked to previous items on a fraction of trials rather than display a small but consistent bias towards previous items. However, the latter has been reported (e.g., in primate spatial working memory, Papadimitriou et al., J Neurophysiol 2014). The ready availability of delayed estimation datasets online (e.g., from Rademaker and colleagues, https://osf.io/jmkc9/) will facilitate in-depth investigation and reconciliation of this issue.

Second, the bulk of the modeling efforts presented here are devoted to a circuit-level description of how putative posterior parietal cortex (PPC) and working-memory (WM) related networks may interact to produce such volatility and biases in memory. This effort is extremely useful because it allows the model to be constrained by neural observations and manipulations in addition to behavior, and the authors begin this line of inquiry here (by showing that the circuit model can account for effects of optogenetic inactivation of rodent PPC). Further experiments, particularly electrophysiology in PPC and WM-related areas, will allow further validation of the circuit model. For example, the model makes the strong prediction that WM-related activity should display 'jumps' to states reflecting previously presented items on some trials. This hypothesis is readily testable using modern high-density recording techniques and single-trial analyses.

Finally, while there has been a refreshing movement away from an overreliance on p-values in recent years (e.g., Amrhein et al., PeerJ 2017), hypothesis testing, when used appropriately, provides the reader with useful information about the amount of variability in experimental datasets. While the excellent visualizations and apparently strong effect sizes in the paper mitigate the need for p-values to an extent, the paucity of statistical analysis does impede interpretation of a number of panels in the paper (e.g., the results for the negatively skewed distribution in 5D, the reliability of the attractive effects in 6a/b for 2- and 3- trials back).