Generalization

When modeling the generalization block, visually distinct and visually similar trials were further separated, depending on whether they were used during training (old) or whether they were new examples of each category. This resulted in four separate predictors (i.e., Novel-Similar, Novel-Distinct, Old-Similar, Old-Distinct), with the data from each subject being separately analyzed at a first-level analysis. Each stimulus onset time was convolved with a hemodynamic response function and entered into a general linear model with their temporal derivatives to estimate beta weights. Contrast maps were created for each subject, showing areas that were differentially engaged during visually similar vs. visually distinct categories. A group analysis was then run using FLAME 1, combining contrast maps across participants. Figure 8 shows regions that were more engaged during visually distinct trials over visually similar trials (red), and vice-versa (blue). Individual voxels were considered to be active when reaching *Z* > 1.8 and surviving a whole-brain cluster size threshold set at *p* < 0.05 [27].

The results from our univariate analysis show that the left caudate nucleus, left superior frontal gyrus, and left inferior frontal gyrus were significantly more engaged on visually similar trials when compared to visually distinct trials (Figure 8). Caudate nucleus, instead of hippocampus, is one of the only observable differences between the training and generalization contrasts for this condition. Table 3 presents a list of the top 11 clusters from this contrast. In addition, the lateral occipital cortex and right fusiform gyrus were engaged significantly more for distinct trials over visually similar trials during generalization. Table 4 shows a summary of the top 11 clusters.

**Figure 8.** Univariate contrasts of visually similar > visually distinct (Red) and visually distinct > visually similar (Blue) during generalization displayed in (**a**) axial, (**b**) medial sagittal, and (**c**) more lateral sagittal views. Red: Frontal control regions were engaged significantly more during the visually similar trials compared to visually distinct trials during generalization. A cluster over caudate nucleus was also found. Blue: Visually distinct trials relied more heavily on lateral occipital cortex compared to trials separable by a counting rule.

**Table 3.** Cluster location and size for similar > distinct contrast in generalization block.


**Table 4.** Cluster location and size for distinct > similar in generalization block.


#### 2.2.3. Multi-Voxel Pattern Analysis

The univariate analyses provided preliminary evidence that participants may engage distinct neurocognitive processes when categorizing visually distinct vs. visually similar trials, albeit at an exploratory threshold. As a second approach, we employed multi-voxel pattern analysis (MVPA; e.g., [29]), which might provide additional sensitivity. Specifically, we asked whether a machine-learning classifier could distinguish, based on the pattern of activation across voxels, into which condition a current trial belonged. First, cortical parcellation and subcortical segmentation was performed while using Freesurfer software for each participant [30,31]. Given prior work on visual memory and the dissociations between rule-based and similarity-based categorization, we focused on regions within the frontoparietal attentional network and the midline, in addition to the posterior visual cortex [32–34]. Specific Freesurfer-defined ROIs included superior parietal lobe, anterior cingulate cortex (ACC), medial orbitofrontal cortex (MOFC), inferior parietal lobe, inferior frontal gyrus (IFG), and fusiform gyrus.

One participants' data were lost between the time we performed the univariate analysis and MVPA due to a site-wide data loss, which left nine out of 11 subjects for MVPA. We modeled the functional data using a separate regressor for each trial to construct a betaseries representing activation patterns that are associated with each individual trial to obtain trial-specific activation patterns for each participant [35]. Each betaseries was smoothed (σ = 3) before being co-registered to each participant's high-resolution anatomical image while using Advanced Neuroimaging Tools (ANTs) toolbox [36]. Data were kept in native the space of each participant for the classification. Classification analysis used a linear Support Vector Machine (SVM), as implemented by the LinearCSVMC classifier in PyMVPA (pymvpa.org). Data from all runs were included to obtain enough training samples for the classifier. A leave-one-run-out crossvalidation was chosen, as it would maximize the amount of data in each training fold for our sample size [37,38]. Within each ROI separately, the classifier was trained on data from five out of six training runs and tested on the left-out run. The classification accuracy was then averaged across the six cross-validation folds. Two binary (pairwise) classifications were performed: visually similar category 1 vs. visually distinct category, and visually similar category 2 vs. visually distinct category. We subsequently averaged their results together, as there were no di fferences in classification accuracy between the two (as expected).

A one-sample *t*-test was used against a baseline value of 0.5 (50% chance for two equally frequent categories). Figure 9 shows the classification analysis in each ROI, indicating dissociable patterns of activation evoked during categorization of stimuli that presumably required rule application vs. those who could be based on perceptual similarity. The IFG ( *M* = 0.66; *t*(8) = 4.23, *p* = 0.003), inferior parietal lobe ( *M* = 0.70; *t*(8) = 3.65, *p* = 0.007), superior parietal lobe ( *M* = 0.76; *t*(8) = 5.8, *p* < 0.001), MOFC ( *M* = 0.58; *t*(8) = 3.3, *p* = 0.011), and fusiform gyrus ( *M* = 0.62; *t*(8) = 3.75, *p* = 0.006) all predicted category membership with statistically significant accuracy. The ACC ( *M* = 0.58) did not reach significance, *t*(8) = 2.02, *p* = 0.078.

**Figure 9.** The inferior frontal gyrus (IFG), inferior parietal cortex, medial orbitofrontal cortex (MOFC), superior parietal cortex, and fusiform gyrus were able to classify between our two conditions with significantly above-chance accuracy. Amongst these regions, the superior and inferior parietal cortices provided the most reliable classification. The Anterior Cingulate Cortex (ACC) did not reach statistical significance.
