**1. Introduction**

Category learning has been a productive paradigm for studying learning and memory and it refers to the development of the ability to group objects belonging to the same category and differentiate objects belonging to different categories [1]. There is not a single mechanism of category learning. Research using category learning models have outlined that there are distinct neural mechanisms associated with different learning stages [2]. Furthermore, we know that different tasks engage dissociable memory systems that are optimized for the type of learning involved—even for seemingly similar tasks, such as in categorization [3–7]. This makes it difficult to uniquely attribute the changes in brain activity to either distinct learning systems or representations of the distinct mechanisms that are associated with different task sets. In other words, the two bodies of literature can complicate comparing brain activity across tasks as subjects between those tasks could either be at different stages of learning or be relying on different categorization strategies that are served by dissociable memory systems.

From the skill acquisition perspective, a succinct model of learning has been proposed that describes reliance upon executive functions, depending on the stage of learning: early or late. Under the dual-stage model, the early stage of learning is marked by being heavily reliant on controlled processes, requiring a person to be actively attentive and dependent on limited working memory capacity. In contrast, the late stage is defined by its lack of reliance on controlled processes, reflected as

automated performance, and it is not limited by working memory capacity and can be subconsciously carried out under the right context [2].

Modern imaging evidence has delineated distinct brain networks that are involved in the two learning stages [8]. The frontal lobe is responsible for the executive monitoring of unfamiliar stimuli; a process that is integral to the early stages of learning. In contrast, cortical regions of the posterior corticolimbic system are engaged when subjects demonstrate proficient performance in the late stages of learning [8,9]. These posterior corticolimbic structures, which include the hippocampus and posterior cingulate cortex (PCC), consolidate information and, with su fficient practice, enable performance to be more automated, thus removing the need for executive control.

Finer details regarding how the brain changes as a person learns to recognize category structures can be understood from the perspective of the dual stage theory [10]. The dual model of sensory information processing is based on evidence that suggests two separable and parallel systems operate on incoming sensory data. The first is a ventral "what" system that is responsible for the identification of an event or object and it includes the sensory specific cortices (such as visual cortex) and the ventral limbic system, which includes the parahippocampal gyrus, piriform, entorhinal cortex, and the amygdala [11,12]. The second processing stream, as exemplified by dorsal cortical regions of the parietal lobe, is referred to as the dorsal or "where" pathway, and it specializes in spatial analysis of stimuli [11,12]. Information from both streams converge at the hippocampus, which is a structure situated in the medial temporal lobe (MTL) that plays a key role in organizing input to link memories by their contextual representation [13]. This feedback structure allows for the hippocampus to organize memory retrieval based o ff "what" occurred or "where" something occurred and makes it an essential mechanism for memory retrieval. With further training, the hippocampus is able to perform declarative recall with less need for controlled attention and input from these two sensory pathways; reflecting the early/late transition outlined in the dual stage model. However, a relevant shortcoming to both the dual processing and dual stage models is that as they exist, they do not account for the evidence of other types of memory systems and their possible di fferential reliance on the brain mechanisms that are described in the models. Another shortcoming is that they do not consider the possibility that di fferent memory systems could be simultaneously engaged during a task, either in competition or working in conjunction, and at di fferent learning stages to optimize learning.

Multiple mechanisms that rely on dissociable memory systems have been implicated in categorization and category learning. One distinction is between strategies that require the application of an explicit rule (rule-based categorization) and those that rely on perceptual similarity (examples: [14,15]). For example, in a family resemblance structure, stimuli that belong to the same category share several common features, with none of them being necessary or su fficient for category membership [16,17]. Categorization relies on the overall similarity rather than an explicit rule. The perceptual similarity system involves posterior visual areas and does not heavily rely on working memory [18–20]. Perceptual similarity allows for making rapid judgements regarding category membership without using much cognitive resources, but falls short in its ability to classify objects when within-category similarity is low or between category similarity is high [21].

In contrast, in rule-based categorization, category membership is dictated by an explicit, verbalizable rule [15]. Rule discovery is commonly achieved through explicit reasoning and hypothesis testing that heavily relies on working memory and selective attention, which are supported by the working memory system in prefrontal cortex and caudate nucleus [22]. The working memory system, within the context of rule-based categorization, allows for participants to focus on individual diagnostic dimensions while ignoring the irrelevant features within the task. This allows for accurate categorization when the within-category variance is high and between-category variance is low. However, rule-based categorization is cognitively expensive and sensitive to distractions when compared to the perceptual similarity system [3,23].

Prior research has focused on creating tasks that exaggerate the preferential recruitment of one system over another to provide compelling evidence for the existence of multiple systems. Evidence from these studies has shown that performance is hindered when the participants fail to engage the memory system optimal for a given category structure. The composition of natural categories contains elements of rule-, *and* perceptual-based systems, suggesting people may be switching between systems within a single task. Identifying signatures of distinct memory systems within single tasks would allow us to better understand how each system contributes to performance and how these systems fit within the expertise development framework.

The main goals of the presented studies were to understand the degree to which distinct learning and memory systems are recruited within the same, real-world task. We implemented a categorization task that was designed to encourage participants to switch between categorization strategies on a trial-by-trial basis and then measured the underlying neural activity in two separate experiments while using functional Magnetic Resonance Imaging (fMRI) and dense-array Electroencephalography (dEEG). The goal of our first experiment, which was a low-sample pilot, was to utilize the spatial resolution of fMRI to establish the overall effectiveness of our task in engaging different memory systems for different trials within the same task. A successful proof-of-concept and spatial distribution in the fMRI pilot then motivated the second, dEEG experiment. In our fully powered dEEG experiment, we studied the time course by which these strategies (and their underlying memory systems) were engaged on a trial-by-trial basis. More specifically, we were interested in understanding when, on a given trial, we can accurately dissociate between verbal and non-verbal rules and the associated memory systems. As the brief involvement of limited attentional and working memory resources may beneficial for the optimization of categorization strategy to the task [24], mapping the timing of the initial convergence and subsequent divergence of distinct categorization processes can provide new insights regarding how distinct systems compete and cooperate to optimize performance. Rough estimates of the anatomical differences between these systems were made while using the EEG data and a novel machine learning approach.

#### **2. fMRI Pilot Experiment**

#### *2.1. Materials and Methods*
