Abstract
Whether probabilistic category learning in the feedback-based weather prediction task (FB-WPT) is mediated by a non-declarative procedural system remains a central question in cognitive psychology. This study compared the effects of training time and verbal working memory load on FB and paired-associate (PA) versions of the WPT. Results from two experiments demonstrated that while PA learning—a declarative process—is significantly impaired by reduced training time and concurrent working memory tasks, FB learning remains robust. These findings provide strong evidence for the multiple-systems theory, suggesting that FB-WPT recruits a procedural learning system independent of the declarative resources required for PA tasks.
Introduction: The Debate Over Category Learning
A fundamental debate in cognitive science concerns whether human category learning is governed by a single declarative system or multiple distinct systems—specifically, a conscious "declarative" system for rules and a subconscious "procedural" system for habits. The Weather Prediction Task (WPT) is a standard tool for investigating this. In the WPT, participants predict an outcome (Rain or Sun) based on combinations of four cards that are probabilistically related to the results.
While early theories posited that feedback-based learning in the WPT is procedural, "single-system" proponents argue it is merely a complex declarative task. This research tests these competing theories by examining if feedback-based learning is resistant to stressors that typically disrupt declarative memory: time pressure and high working memory load.
Methods and Experimental Design
The researchers conducted two experiments comparing two learning conditions:
- Feedback-Based (FB): Participants predict an outcome and receive immediate feedback (e.g., a smiley face for correct guesses).
- Paired-Associate (PA): Participants view the cards and the outcome simultaneously, requiring them to memorize the association.
Experiment 1: Training Time Manipulation Participants were given either 7 seconds or 3 seconds per trial. Declarative systems generally require more time for elaborative encoding, whereas procedural systems are expected to operate more automatically.
Experiment 2: Working Memory Load A secondary "Sternberg memory-scanning task" was introduced to consume verbal working memory resources. Participants had to maintain a four-digit string in memory while performing the WPT.
Learning was assessed through Optimal Accuracy (choosing the statistically superior outcome), Task Knowledge (explicit probability ratings), and Self-Insight (reported reliance on specific cards).
Key Results and Insights
The study revealed a clear dissociation between the two learning conditions:
- Robustness of Feedback-Based Learning: In Experiment 1, reducing training time significantly lowered accuracy in the PA condition (from ~80% to ~72%) but had no significant effect on the FB condition (~80% accuracy in both). Similarly, in Experiment 2, the working memory load impaired PA performance but left FB accuracy unaffected (~82%).
- Dissociation of Knowledge Types: Interestingly, while the secondary task in Experiment 2 did not hurt FB accuracy, it did reduce participants' explicit knowledge of the task. This suggests that participants in the FB condition were learning the categories procedurally—"knowing" the answer without being able to consciously describe the underlying rules.
- Support for Multiple Systems: These results suggest that the FB task relies on the basal ganglia and dopamine-mediated reward signals. This procedural system can strengthen synaptic connections automatically upon receiving feedback, functioning independently of the conscious attention required by the declarative system used in PA learning.
Discussion and Future Directions
The findings strongly support the multiple-systems theory by showing that FB-WPT is mediated by a procedural system that is resilient to cognitive load. However, some limitations remain. Discrepancies with previous studies using different secondary tasks (like the Numerical Stroop task) suggest that different types of cognitive load (visual vs. verbal) may interfere with learning systems differently.
Future research should incorporate neural verification, such as fMRI or studies with clinical populations (e.g., patients with Parkinson’s disease), to confirm the involvement of the basal ganglia. Additionally, computational modeling could help determine if participants shift strategies (e.g., from multi-cue to single-cue) when under high cognitive load.
Summary updated