Acquisition curves (thick lines, approximated by a Savitzky-Golay filter) for mice trained with different temporal arrangements of CS^{+}/CS^{−} similarity (SSIM). Thin dashed lines correspond to chance level at 50% correct choices. (A) Training with equiprobable CS^{−} stimuli, linearly sorted into increasing (blue) or decreasing (red) SSIM values from a ‘narrow’ SSIM range. (C) Training with the same stimuli as in (A) arranged into increasing (blue) or decreasing (red) gradients of inter-training unit similarity (ΔSSIM); this oscillating arrangement maximizes similarity gradients, whereas linear sorting minimizes them. Asterisks below charts depict above chance choice values (Wilcoxon test, **P* < 0.05). The average performance at 150 ± 5 training units with identical average SSIM was: SSIM_{inc,narrow}: 69% ± 5%, SSIM_{dec,narrow}: 61% ± 6%, ΔSSIM_{inc,narrow}: 82% ± 3%, ΔSSIM_{dec,narrow}: 65% ± 3%; one-way ANOVA, *F*_{3,35} = 9.44, *P* = 0.02. (B, D) Correct choice probability against average stimulus similarity for the corresponding trials from each group. Pooling the data from the 6 groups trained with varying similarity rendered a linear regression with a clear negative slope, an adjusted R^{2} = 0.99, and an overall similarity discrimination threshold of SSIM ≈ 0.33 (not illustrated), and shuffling the SSIM data by random permutations rendered infinite slopes (not illustrated). (E–G) Average performance of last 30 training units at the end of training. Asterisks indicate above chance choice values (Wilcoxon test, **P* < 0.05, ***P* < 0.01) and lowercase letters depict the differences between groups (paired-comparisons, one-way ANOVA, followed by Bonferroni's post-tests, *P* < 0.05). Number of mice per group in parentheses.

## PubMed Commons