Restoring spatial working-memory reconstructions

Status: main prose and supplementary Methods read; figure/equation/table audit incomplete at the user's instruction to finish. This is a bounded reading note, not a completed full-source audit. Updated 2026-09-22.

Sprague, T. C., Ester, E. F., & Serences, J. T. Restoring Latent Visual Working Memory Representations in Human Cortex. Neuron 91, 694–707. Published 3 August 2016. DOI; PubMed. Peer-reviewed human experiment.

Factual core — concise source paraphrase

Six participants remembered one or two briefly presented colored locations. After an eight-second delay, a fully valid color cue could select one of two remembered items; a second eight-second delay preceded a position report. Valid cues reduced final recall error relative to neutral cues, though performance remained worse than remembering one item throughout. An inverted encoding model, fitted on independent spatial mapping runs, reconstructed remembered locations from visual and parietal fMRI patterns. Its fixed transformation yielded increased target-aligned reconstruction fidelity after valid cues in several regions. The improvement primarily involved amplitude rather than consistent narrowing. Better final recall accompanied larger reconstruction amplitude in the combined-region analysis. The authors interpret recovery as recruitment of information held in a latent state. The observations establish flexible access to spatially specific neural representations; they do not directly measure a silent substrate or establish that this reconstruction metric orders the information in the complete underlying memory state.

Task and analysis audit

• Sample and burden. Six adults, five female, aged 22–29, completed three approximately two-hour main scanning sessions, separate retinotopic mapping, and two to four behavioral practice sessions. Each contributed 324–378 main trials. No participant exclusion was reported in the read Methods. A participant's earlier experience with a similar task is acknowledged; an exclusion sensitivity claim is reported without the corresponding data. This is intensive within-person calibration, not a six-trial or calibration-free demonstration.
• Information available to the cue. Two dots appeared for 500 ms. An initial cue selected one item or instructed retention of both. Eight seconds later, the valid retro-cue selected the red or blue item with 100% validity. It supplied relevance and color identity, but no coordinate. Locations were continuously jittered around positions on a ring; the response tested a randomly selected horizontal or vertical coordinate, unknown during maintenance. Thus the experiment goes beyond category-only decoding while still providing known stimuli, task structure and selection cues to the analyst.
• Behavioral endpoint. Recall after the complete delay improved for valid versus neutral cues, and remained worse than the remember-one condition. There was no behavioral report immediately before the retro-cue. Consequently, the behavior does not directly show restoration above the person's own pre-cue precision: protection against further loss and changes in selection or reporting remain possible contributors.
• Training and held-out unit. The spatial model used 37 prescribed spatial basis functions. Voxel weights were fitted separately by participant, session and region using the four independent mapping runs of that session, then held fixed across main-task conditions and time points. This is useful transfer from an independent mapping task to memory-delay activity, not cross-person generalization. The reported mean representations were rotated and translated using the true target coordinates; this is a truth-aligned evaluation display, not a demonstration of a blinded recipient recovering an unknown coordinate from that averaged display alone.
• Metric and comparisons. The reported fidelity projects an angular reconstruction onto the direction of the true target: conceptually, the mean of reconstruction activation times the cosine of angular offset. It depends on amplitude as well as shape. It is not a calibrated posterior, a mutual-information estimate, or a direct measure of single-trial spatial precision. The primary comparison of second- versus first-delay fidelity was positive in V1, IPS0, IPS1 and the combined-region reconstruction. The broader time-pair analysis was supplementary and is only partly audited here.
• Reconstruction parameters. Recovery was chiefly in fitted amplitude; changes in fitted size were inconsistent. The post-cue amplitude could exceed remember-one amplitude despite poorer behavioral accuracy, cautioning against a one-to-one identification of this signal with retained precision. Exact equations and all parameter tables still require visual/numeric verification; no algebraic formula beyond the conceptual fidelity definition is certified in this note.
• Timing and data processing. Delay estimates used BOLD measurements at 6.75–9.00 seconds and 15.75–18.00 seconds after trial onset. These are hemodynamic observations, not instantaneous readings of neural content at the cue. Mapping and memory runs were separate, but observations were still within the same people and sessions. In reconstruction analyses, some 180-degree-separation trials were dropped to balance the other item's relative locations.
• Uncertainty and behavioral association. The described bootstrap resampled individual trials pooled across participants and sessions; it did not describe a participant-cluster bootstrap for population generalization. Trials were divided into lower- and higher-error halves within participant, session and condition before comparing reconstruction parameters. The combined-region amplitude association is retrospective; it is not prospective prediction, an independent validation cohort, or evidence that amplitude mediates the cue's effect. Several individual-region associations did not survive the stated multiple-comparison correction.

Interpretation and extraction implications — original reasoning

The positive result should be retained: a spatial model learned in a separate task tracks remembered within-category locations, and the same readout changes with relevance and later performance. This is not merely a language prior producing plausible text, nor merely discrimination of broad stimulus categories. The cue does not contain the missing coordinates. However, the study did not quantify how much new, unknown detail a recipient could recover conditional on all other released information.

The authors' information-preservation argument needs an observation-model qualification. A lower-dimensional readout may become stronger even if the full neural state never lost the relevant information. Increases in a selected reconstruction metric do not by themselves establish increases in mutual information in the complete population state. A fixed decoder prevents explanations based solely on refitting the decoder, but does not eliminate changes in neural gain, accessibility, interference or measurement noise. Those are alternatives to test, not established explanations of this dataset. Nor does the study directly locate content in synaptic weights, rule out ongoing activity outside the measured projection, or demonstrate transfer of an autobiographical episode to another system.

For an eventual extraction experiment, freeze the allowed context, calibration and cue inputs; hold out exact target coordinates from the recipient; score blind probabilistic coordinate estimates with a proper score; and report both error and calibration. Compare valid cue, neutral cue and suitable context-only baselines with matched observation opportunities. Separately test whether a cue protects subsequent retention or restores accuracy relative to a measured pre-cue state, accounting for the fact that measuring that state can itself alter memory. Keep this proposal distinct from what the present paper did.

The team is separately reading Schneegans and Bays (2017), which offers an observation-model challenge. Its simulations are not part of this note's primary-source reading, and no simulation result is attributed to Sprague's participants.

Sources and exact reading boundary

• Published-layout main PDF, 15 PDF pages: one highlights cover plus 14 journal pages; text. Obtained through the author laboratory publication index and its public PDF link.
• Official supplement PDF, 27 pages; text. Publisher source.
• Publisher Excel Table S7 and Table S8; S7 source, S8 source. ZIP/workbook identities verified; cells not substantively read.
• Provenance and SHA256 manifest covers all six source/text artifacts. Public accessibility does not establish an unrestricted reuse license.

Read: all main prose through Discussion and Experimental Procedures; relevant main captions as encountered in extracted text; supplementary Methods on PDF pages 20–26; captions S1–S4 and S6–S8. Visually inspected main Figures 1 and 2 only, including their captions. The S5 caption remains partially unread because text extraction interleaved labels and the returned output was truncated.

Not completed: visual inspection of main Figures 3–8, all supplementary figures and equations; systematic reading/verification of Tables S1–S6; Table S7/S8 numeric cells and page-19 captions; complete version/correction audit. References were retained, not read as additional primary papers. No raw data, empirical code execution, model reproduction, new experiment, or participant access occurred. The user instructed the research team to finish before the remaining checks, so this entry must remain partially audited rather than being promoted to full-main-plus-supplement verification.