Devraj et al. (Psychonomic Bulletin Review. doi:10.3758/s13423-023-02448-2, 2024) argued that findings which suggest that memories for items become less accessible over time conflict with categorisation findings where exemplar performance improves across training. To reconcile this, they highlighted that under real-world conditions items tend to reappear less frequently over time, thus preferentially maintaining new items can improve performance. Typical categorisation experiments instead distribute exemplars uniformly across trials. Instead, under a power-law stimulus distribution Devraj et al. showed worsening fit for exemplar classification models across trials. They used this as evidence that forgetting behaviour adapted to task demands, reducing exemplar were accessibility and encouraging prototype use for classification decisions. By re-analysing the same data, we argue instead that this pattern can be produced with exemplar-forgetting in both conditions. By systematically increasing in the delays across which stimuli were tested, their Experimental condition exaggerated the effects of forgetting on performance in later trials compared to the Control condition. This resulted in a reversal of performance growth across trials – instead leading to a steady decline in performance. As exemplar model-fit advantage is expected to vary with performance, we suggest that trends in this advantage are not diagnostic of a shift in classification strategy. We found that a forgetting-function improved exemplar model fit to Devraj et al.’s data, and under reasonable parameters could predict the observed patterns of performance and model-fit a-priori. Compared with a strategy-shifting mixture model, exemplar-forgetting provided equivalent fits despite being more theoretically parsimonious. We suggest power-law memory decay does not produce a tension between categorisation and memory findings, as increased forgetting is found across longer retention intervals whereas the delay between exemplar learning and classification remains constant across typical categorisation experiments.
Hutchinson et al. (Mon,) studied this question.