Recalling a single object: going beyond the capacity debate

Rakesh Sengupta, Christelle Maria Lewis, Raju S. BapibioRxiv 2020 · DOI

Abstract

Working memory (WM) is traditionally viewed as a limited-capacity system, with debates centered on whether it consists of discrete "slots" or a continuous "resource." Both models imply that recalling a single object should be near-perfect. This study challenges this assumption by investigating recall accuracy for multiple features of a single object across three experiments. Results demonstrate a significant "recall penalty": errors for color, location, and size increase with the serial position of the question. This suggests that the act of retrieval itself disrupts working memory representations, a finding that necessitates a shift from passive storage models to more dynamic, process-oriented frameworks.

Introduction: The Capacity Debate

For decades, the "Capacity Debate" has dominated visual working memory (VWM) research. Scientists generally fall into two camps: the Slot Model (we have a fixed number of "slots" for items) and the Resource Model (we have a limited pool of "neural fuel" to spread across items). Both models share a common assumption: if you are only asked to remember a single object, your performance should be near-perfect. Since one object doesn't fill all the slots or deplete the resource pool, there should be no "crowding" in the mind. However, this paper challenges that assumption, asking whether the actual process of recalling information changes the memory itself.

Methods and Experimental Design

The researchers conducted three experiments using a simple stimulus: a single, colored, oriented line of a specific size, presented briefly on a screen.

Experimental procedure showing the stimulus presentation and the randomized feature recall questions.
  • Experiment 1: Participants (n=20) saw the line for 1000ms and were then asked 1 to 4 questions in random order about its features: Color, Location, Orientation, and Size. Responses were recorded on a continuous scale.
  • Experiment 2: To rule out "uncertainty" regarding the number of questions, the researchers fixed the number of probes to exactly two.
  • Experiment 3: To rule out simple "time decay" (forgetting because time passed), they varied the delay before a single recall question (1 second vs. 2 seconds).

The team used z-scores to normalize errors across different units (e.g., degrees for color vs. pixels for location) and applied repeated measures ANOVA to look for significant trends.

Key Results: The Recall Penalty

The primary finding across Experiments 1 and 2 was a significant Serial Position Effect. As more features were recalled, the accuracy of subsequent recalls declined.

Recall error (z-score) as a function of question order for different features in Experiment 1.

In Experiment 1, the main effect of question position was highly significant ($F(3, 228) = 48.79, p < 0.001$). Post-hoc tests showed that for Color, Location, and Size, errors jumped significantly between the 1st and 2nd positions. Interestingly, Orientation recall remained stable regardless of whether it was asked first or last, suggesting it might be encoded more robustly.

Experiment 3 showed that a 1-2 second delay didn't significantly increase errors. This proves the drop in performance in the other experiments wasn't just due to the memory fading over time; it was specifically triggered by the act of responding to previous questions. Furthermore, the researchers found mostly positive or no correlations between feature errors, contradicting the idea of a simple "resource trade-off" where subjects prioritize one feature over another.

Discussion and Implications

These findings suggest that working memory is not a passive container. Instead, the "executive" process of retrieval appears to disrupt the remaining stored information. The stability of orientation recall suggests that different features may be processed in distinct neural pathways—perhaps early visual areas (V1/V2) for orientation versus higher-order areas for color.

This "recall penalty" poses a challenge to both slot and resource models, as neither predicts such a sharp decline in performance for a single object within a short timeframe. Future work should explore the neural mechanisms of this interference, potentially linking it to "inhibition-of-return"—a phenomenon where the brain suppresses recently processed information to focus on something new.

Summary updated

Abstract

Working memory is now established as a limited capacity system. The debate regarding working memory has been largely between slots and resource based models. The resource model suggests that as the number of items increases, precision of recall decreases because neural resources are dynamically allocate to all the objects needed for task. Slot model on the other hand implies that an item is stored either with the highest precision or not at all. If both these models stand true then quality of memory performance would be near perfect for a single object. However, that may not be the case. In the current work, we investigated recall accuracy for feature(s) of a single object in three successive experiments. In all three experiments, the memory array consisted of a single colored oriented short line presented a short distance away from the center of the display for 1 sec. We probed recall of features from the set of color, location, orientation, and size. In experiment 1 number of recall question varied between 1 – 4 with the order randomized in each trial. In experiment 2 we chose to probe only two feature recall questions, whereas only one recall question was probed for the third experiment. In experiment 3 we varied the delay before the recall probe between 1 and 2 s. The recall response for each feature was mapped on to a continuous variable. Subjects used a color wheel to respond to color, on-screen mouse click to indicate center of the line location, click away from the center to indicate size, and a mouse click to the periphery of centered circle to indicate orientation with the slant of the resulting radial line. We calculated z-scores of errors for each feature for every subject separately. In experiment 1 that for color, location, and size the errors increase significantly with the position of the questions asked. In experiment 2, the errors increased significantly between questions for color and location (but not for orientation and size). In experiment 3, we did not see any significant increase in error with recall probe delay. Overall run-time for each trial was within 10 secs, well within the limits of operation of working memory. This drop in performance poses questions for memory mechanisms proposed by slot and resource models as both would predict near-perfect recall within the time-period for the trials.
BibTeX
@misc{sengupta2020recalling,
  author = {Rakesh Sengupta and Christelle Maria Lewis and Raju S. Bapi},
  title = {{Recalling a single object: going beyond the capacity debate}},
  year = {2020},
  doi = {10.1101/2020.02.17.951921},
  howpublished = {bioRxiv},
}
Figure