*2.2. Results*

#### 2.2.1. Behavioral Results

Data from one participant were excluded due to noise caused by motion during scanning, leaving 10 out of the 11 subjects for analysis. Figure 4 shows the average classification accuracy for each category across the six training runs. Visual inspection indicates that participants learned the visually distinct category faster when compared to the two visually similar categories, but by the end of run 4 they were able to accurately identify members of all categories.

**Figure 4.** Subjects accurately categorized the visually distinct category quicker than the two visually similar categories. Accuracy for the visually similar categories peaked between runs 4 and 5, which we infer is the time at which subjects discovered the counting rule.

A confusion matrix shows that subjects commonly mixed up the two visually similar categories when making errors, and rarely mixed up the visually distinct category with any other. By block 4, the subjects limited their confusion, which is indexed by the mostly uniformly colored bars in Figure 5. We can infer that this was the point at which most subjects discovered the explicit counting rule that allowed for them to di fferentiate between members of the two visually similar categories (Figure 5).

**Figure 5.** Visualization of the confusion matrix during classification. During the first 3 training blocks, subjects commonly confused the two visually similar categories for one another. By run 4, subjects were able to accurately dissociate between these two categories. Subjects rarely confused any other category when classifying formations in the visually distinct category.

The generalization run was included to test for the hallmark of category knowledge—the ability to generalize category labels to novel category examples. On average, the subjects completed the generalization run with 92% accuracy for the visually distinct category and 88% accuracy for the visually similar categories (Figure 6). Had subjects been relying on the declarative recall of individual stimuli throughout training, their performance in the generalization run would have been closer to chance (33%).

**Figure 6.** Categorization accuracy in the generalization block.

In the post-test questionnaire, nine out of 10 participants indicated that they used a counting strategy to distinguish between the visually distinct categories, meaning that they accurately identified the defining number of players on the line of scrimmage separating the two categories. The same participants reported relying on perceptual similarity to identify formations in the visually distinct category. This indicates that participants treated the stimuli as expected, using an explicit rule specifically when between-category similarity was high as compared to the within-category similarity. The remaining one participant reported using declarative recall for all stimuli, whereby they memorized each formation individually instead of relying on the intended counting rule. This participant categorized new stimuli in the visually similar categories with an accuracy near chance during the generalization block, which indicated that relying on declarative recall instead of discovering the counting rule was ineffective for generalization in this task.

#### 2.2.2. Univariate fMRI Analysis
