Reading focus
Syllabus: How Do We Learn What Predicts What?
This selection is Chapter 3 of Classical Conditioning II, pp. 64–99. Translate the equations into changes in prediction and follow how the same rule handles different compound-training arrangements.
Why you’re reading it. Move from the contingency finding to a specified process that makes testable predictions. Understanding the argument does not depend on reproducing every numerical calculation.
A question to consider. After A already predicts an outcome, X is added and AX receives the same outcome. Why does the summed prediction matter for learning about X, and what would changing the outcome let you test?
Why this reading
This chapter turns a verbal idea about surprise into a model with explicit predictions. Its central move is to make learning about one cue depend on the combined prediction supplied by the cues present, allowing the same rule to address several forms of cue competition.
Before you read
A conditioned stimulus (CS) is a cue; the unconditioned stimulus (US) is the event it signals. Associative strength is a model quantity, not a directly observed response. Prediction error is the difference between the outcome supported by the US and the current combined prediction. The learning-rate parameters regulate the size of an update.
Read for
- Translate the update rule into words and distinguish a cue’s own associative strength from the compound’s summed strength.
- Trace why prior learning about A can reduce learning about X when AX is followed by the same outcome, and why a larger-than-expected outcome changes the prediction.
- Notice that even a reinforced trial can reduce associative strength when the combined prediction exceeds the outcome; examine the assumptions needed to connect strength with behavior.
Consider afterwards
- Two separately trained cues each predict one outcome. What does the model predict when they are combined but the outcome remains unchanged? State which assumptions your answer needs.
Editorial context
The chapter proposes a formal account, not a complete description of learning or a demonstrated neural mechanism. It leaves parts of the mapping from associative strength to responding unspecified. Soto and colleagues’ later appraisal discusses its continuing influence alongside its limitations. Explaining a familiar effect is one achievement; specifying conditions under which the effect should change is another.