4.2.1. ERPs

The MFN amplitude in this experiment was larger for the visually distinct category, albeit the significance of this e ffect was only marginal. The moderate di fference in amplitude between our categories support recent findings that sugges<sup>t</sup> multiple categorization strategies—even those inferred to rely very little on the working memory system—need executive functions in order to select the memory system that is optimal for a task [50]. For stimuli that would benefit most from perceptual similarity, this requirement of e ffortful control would be very brief—commencing well before an action is committed [50]. The latency of the MFN (180–300 ms) corresponds to the initial orienting of attention in a visuomotor association task and, thus, we propose that the MFN in our task is indexing the controlled attention that is required to select the memory system best suited for categorizing the presented stimulus and does not depend on the optimal system needed to perform a task.

Like the MFN, the P3b in our experiment was larger for the visually distinct stimuli when compared to stimuli that required the application of an explicit rule. Our initial assumption for this component was that the amplitude should be largest for the visually similar categories based on the perceptual similarity literature, which typically describes the robust activation of posterior visual cortex (and no posterior corticolimbic areas) for the visually distinct category [18–20]. However, this would only be the case if the participants were exclusively relying on perceptual similarity to categorize members of the visually distinct category. High trial counts could result in subjects utilizing di fferent systems to categorize formations in the visually distinct group as their performance improves. When we analyzed the strategies that subjects were using post-hoc, 89% of participants reported using an explicit counting rule or declarative recall for categorizing the visually distinct formations (68% of the total count being declarative recall and 21% counting rules), while only 11% reported using a perceptual similarity strategy. This theory satisfies the two-stage learning *and* multiple memory systems models, where the early stages of learning are marked by a reliance on a variety of strategies (that may rely on dissociable neural systems) to work toward a more routinized and automatic recall of declarative information. However, more studies are required that track changes in the P3b across training to further associate the amplitude of the P3b with specific categorization strategies. Theoretically, we would see changes in the P3b amplitude as participants progress throughout training and, in turn, that should mirror any changes in the strategy they were using for specific categories.

The amplitude of the left LIAN was the largest for the visually distinct condition, whereas the right LIAN was largest for the visually similar condition, although the latter e ffect did not reach statistical significance. The left/right conditional flip makes the interpretation of this component fairly di fficult. At this time, we are unsure whether both components are interpretable on their own, or if the LIAN is a hemisphere-specific component and the e ffect observed on the contralateral side is a byproduct of volume conduction. Luu et al. (2007) found that the amplitude of the right LIAN decreased as subjects acquired the ability to perform spatial analyses in a visuomotor association task, but the amplitude of the component remained unchanged when the targets in the task were digits that evoked the phonological loop [39]. They also found that the amplitude of the left LIAN increased as the subjects acquired digit targets in their task, whereas the amplitude remained unchanged as they acquired the ability to perform spatial analyses. Motivated by the findings of their experiment, we drew an initial assumption that the amplitude of both the left and right LIAN should be largest for visually similar condition in the current experiment. As similarly discussed in our interpretation of the P3b, however, this would only be the case if the subjects exclusively relied on perceptual similarity analyses to categorize formations in the visually distinct category—similar to the spatial analyses that were performed in Luu et al. (2007) [39].

Given the vast majority of subjects in our experiment used rote learning to categorize the visually distinct condition instead of the hypothesized perceptual similarity, one interpretation of our findings is to view them as a contrast between declarative recall of individual stimuli (visually distinct category) and explicit rule application (visually similar categories). When viewed from this perspective, the location of the LIAN coincides with structures that are essential for both forms of analysis, such as the temporal lobe and inferior frontal gyrus (IFG) [64,65]. Based on the higher accuracy for the visually distinct category, our findings that the right LIAN was smaller for this category is in-line with meta-analytic findings that show a right hemisphere-specific reduction in anterior temporal and IFG activity with the development of expertise in visuomotor tasks [8]. We could be seeing right hemisphere-specific reductions in the attentional resources that are needed to categorize the visually distinct group of formations simply because our subjects are consistently at a more advanced stage of learning for this condition when compared to the visually similar condition. Our left LIAN results also become more interpretable through this lens. If our subjects are significantly more advanced at declaratively recalling the visually distinct formations, then we would expect the left LIAN to be larger for this condition based on the findings of Luu et al. (2007) [39]. The amplitude of the left LIAN linearly increased for digit targets in their visuomotor learning task, which theoretically engage the same explicit form of memory as both conditions in our experiment. Thus, the left LIAN differences seen in our study could be reflecting differences in expertise between our subject's ability to categorize the visually similar and visually distinct categories.

#### 4.2.2. dEEG Machine Learning

Using machine learning, we were able to successfully dissociate between our two conditions when utilizing raw voltages distributed across the entire scalp. We were especially interested in the timepoint-by-timepoint classification to identify the earliest point at which we can differentiate between our conditions as subjects view a stimulus. In our study, the onset of a stable period of reliable dissociation was around 200 ms after stimulus onset, which coincides with the initial onset of the MFN ERP component. We interpret this early classification timepoint as reflecting the initial controlled attention required to select a memory system based on the stimulus being presented.

We ran a second machine learning analysis on only the voltages of single groups of electrodes in 20 ms intervals to understand which individual regions were driving the classification accuracy. Our results from this analysis showed that the medial prefrontal, left frontal, and posterior parietal regions collectively contributed to the earliest reliable classification point. fMRI studies using multi-voxel pattern analysis (MVPA) have consistently demonstrated that individual rules can be reliably decoded in frontal and parietal regions [32–34]. Our EEG decoding results expand on these findings by specifying that the pattern representations of these concepts coincide with the initial orientation of attention. Through sufficient trial and error learning, the context under which an action is learned in a visuomotor task becomes tied to each individual stimulus in the task [66]. We can assume that the initial conscious registration of a stimulus prompted a conditioned re-establishment of the explicit rules (the learning context) that would dictate their subsequent action selection since we only analyzed trials after our subjects had been sufficiently trained on the task. This theory could explain why the first pattern dissociation between our two categories happens around the earliest time that a person can explicitly orient attention.

#### *4.3. Category Learning Strategies as a Function of Expertise*

The theories of categorization that formed the basis for our experiments commonly discuss these memory systems individually. However, the results from our experiments indicate that multiple memory systems may develop alongside one another in a single task, alternating from trial-to-trial to meet task demands. The development of expertise within each system could happen independently, and they likely share the same end-goal of automating the attention process with extended training.

Palmeri (1997) made one of the first attempts at describing the time that it takes subjects to reach automaticity while using perceptual similarity versus rule-based categorization [67]. In one experiment, Palmeri had subjects categorize objects with high within-category similarity, whereas in a separate experiment had subjects categorize objects with high between-category similarity, which required the discovery of a rule. The results from these experiments demonstrated that subjects utilizing perceptual similarity reached automaticity notably faster than those that relied on rules. This led to

the development of a new theory termed Exemplar-Based Random Walk (EBRW) which proposes that, when a probe is presented, exemplars stored in memory race to be retrieved with a speed that is proportional to their similarity to the probe. Each one of the retrieved exemplars drives a random walk until su fficient evidence is presented. Once enough evidence has been retrieved, a subject makes a response [68,69].

Computational models of EBRW allow for the reaction times to be sped up by increasing within-category similarity and increasing the number of exposures to an exemplar [67]. This would result in a shorter training period before subjects reach automaticity when categorizing visually similar exemplars. The model also accounts for a longer training period when the subjects are forced to rely more on the random walks or the evidence-gathering aspect of the process when the categories have low within-category similarity and/or high between-category similarity, which was the case for our visually similar categories. EBRW, when interpreted on a conceptual level, helps to explain how implicit and explicit forms of categorization are a simple function of expertise development. The di fferent strategies are called upon, depending on the structure of a category being presented and they share the common function of serving as an intermediate strategy before transitioning to an automatic mode of operation. However, a potential shortcoming of EBRW is that it postulates a single, unitary memory system underlying performance, which does not align well with neuroscience evidence in favor of multiple category learning systems [15,70]. We propose that this theory be altered to accept these processes as the work of distinct memory systems. It is clear that future work is needed to develop new theories for how these distinct systems develop under learning conditions that may require more than one type of system to optimize performance.

#### *4.4. Alternative Interpretations and Limitations*

While we interpret the neurophysiological di fferences between categories to reflect the use of di fferent categorization strategies, a key challenge to clear interpretation is that the conditions di ffer in di fficulty. The subjects had an easier time recognizing and categorizing the visually distinct category, whereas it took longer to do the same for the visually similar categories. The current task and result can be alternatively framed in terms of the di fferences in the relative contribution of top down vs. bottom up processes during learning. Specifically, for the visually distinct categories, subjects could largely rely on bottom-up (stimulus-driven) signals. In contrast, the categorization of visually similar categories requires a greater involvement of top-down signal guiding attention to relevant details to implement an explicit counting rule. Relatedly, we can view our results from a general cognitive resources framework. As stated earlier, the two visually similar categories have a relatively small between-category variance, which would require more working memory resources to discern, and arguably engage, the rule-based categorization system. On the other hand, the between-category variance between each of the two visually similar categories as compared with the visually distinct category is much higher. In theory, making a distinction with high between-category variance should not be as cognitively taxing. The di fferences seen in our ERPs reflect the di fferential allocation of cognitive resources, and this di fference has been argued to be controlled by dissociable memory systems [15,54].

Unfortunatley, a fundamental feature of naturalistic learning environments is that some deviation in individual learning strategy is expected. Although we make the argumen<sup>t</sup> that rule-based and perceptual simliarity-based judgements play an intermediate role on the path to declaritive recall and automazation, there is no clear way to determine whether the subjects switched their strategies with extended training in the current experiments. A future experiment is necessary to further explore the finer details of any inferred stategy shifts related to expertise.
