Reading focus
Syllabus: How Do We Learn What Predicts What?
Follow the argument from observed gradients to psychological distance and the proposed exponential relationship. Keep a mathematical derivation distinct from evidence for its assumptions.
Why you’re reading it. Extend the question of representation beyond compounds to the measurement and prediction of similarity-based transfer.
A question to consider. How could you use an independent similarity task to make a generalization prediction more informative than drawing a curve through the observed test responses?
Why this reading
Generalization requires a way to describe similarity. Shepard proposes a quantitative regularity in psychological space and asks whether diverse results can be understood through a common principle rather than a separate law for each physical stimulus dimension.
Before you read
A generalization gradient describes how responding changes as test stimuli differ from a trained stimulus. Psychological distance need not match physical distance. An exponential decrease falls rapidly at first and then more gradually. A metric specifies how differences across dimensions combine into distance.
Read for
- Separate the observed response gradients from the psychological-space construction used to compare them.
- Examine why the proposed regularity concerns psychological rather than raw physical distance, and how the relations among dimensions affect the metric.
- Distinguish the empirical approximation, its mathematical rationale, and the evolutionary interpretation offered for it.
Consider afterwards
- Two tones equally far from a trained tone in hertz produce different responses. Would this alone refute the proposed law? Explain what independent evidence about psychological distance would be useful.
Editorial context
The title expresses a theoretical ambition, not a guarantee that every generalization task follows the same curve. The initial derivation concerns idealized, immediate generalization after one learning trial; extended training and discrimination raise further issues. Assess whether similarity is constrained independently enough for the proposed law to make a risky prediction, rather than merely redescribe observed responses.