Brain Mechanisms of Concept Learning

The dynamic formation of concept representations during category learning

Concept learning is rapid and flexible; we can adapt newly-learned knowledge to novel situations or changing goals with seemingly little effort. Characterizing the neural mechanisms that support such rapid conceptual learning, especially learning in the face of evolving learning goals, is a critical focus of ongoing research. One recent study (Mack et al., 2016) addressed this topic with an approach combining fMRI and computational modeling to identify the neural machinery of new concept formation. Motivated by human and animal work demonstrating attentional and contextual influences on representations in the hippocampus (Fenton et al., 2010; Aly and Turk-Browne, 2016), this study targeted the formation of new concepts in hippocampal activation patterns. Across two learning tasks, participants learned to categorize multidimensional visual objects according to different category structures: a simple rule based only on a single feature dimension and a more complicated rule based on two feature dimensions. Critically, the visual objects were constant across the tasks, but the underlying organization of the objects into categories differed. This approach required participants to attend to different features between tasks and build new conceptual representations that best matched the underlying category structure.

To investigate the nature of the representations each participant learned in the two tasks, Mack et al. (2016) used SUSTAIN, a computational model that accounts well for behavioral responses during category learning (Love et al., 2004; Love and Gureckis, 2007). Fitting this model separately to each participant's learning behavior provided predictions about how category items were represented in a multidimensional space, and thus how similar they may be perceived by the participant, across the two tasks. The logic followed such that the more similar any two items were represented by the model, the more similar should be their neural activation patterns. To test the role of the hippocampus in building flexible concept knowledge, the model-based predictions were used to interrogate the structure of activation patterns in the hippocampus with representational similarity analysis (Kriegeskorte et al., 2008). If the hippocampus plays a role in building flexible concept knowledge, SUSTAIN′s predictions of category structure should be evident in hippocampal representations. Indeed, a region in anterior hippocampus showed neural representations that were consistent with the model-based similarity matrices. In other words, as learning goals changed, hippocampal representations reorganized to reflect the diagnostic information important for the current task.

Although hippocampal functions offer a powerful toolset for building new concepts (Kumaran et al., 2009; Schapiro et al., 2017), the hippocampus does not act alone. In particular, both rodent (Place et al., 2016; Guise and Shapiro, 2017) and human (van Kesteren et al., 2012; Zeithamova et al., 2012; Schlichting and Preston, 2016) findings point to a functional alliance between hippocampus and medial PFC when encoding new information that overlaps with prior experiences (Preston and Eichenbaum, 2013). Consistent with this notion, Mack et al. (2016) found that anterior hippocampus demonstrated a strong functional coupling with ventromedial PFC (vmPFC) during early learning when concept updating is most potent. This region of vmPFC may, in fact, play an important role itself in guiding attentional tuning during new learning. Neural representations in vmPFC throughout new concept learning have recently been shown to systematically track the efficient mapping of stimuli to categories in a manner consistent with highlighting information that matters and down-weighting irrelevant features (Mack et al., 2019).

Specific and generalized representations supporting categorization

What type or types of memory representations are formed during concept learning and used in categorization decisions has been the focus of cognitive categorization research for decades (Posner and Keele, 1968; Homa, 1973; Medin and Schaffer, 1978; Nosofsky, 1986). Combining neuroimaging with formal models of categorization made it possible to start answering such representational questions in neuroscience research. A study by Mack et al. (2013) tested the degree to which brain signals are consistent with the exemplar model (assuming that categories are represented by the specific category examples one encountered) and the prototype model of categorization (assuming that categories are represented by generalized category representations, or prototypes, abstracted across exemplars). These models are schematically depicted in Figure 1. They identified a network of regions representing specific exemplars during generalization judgments but found no evidence of generalized prototype representations. The categories tested by Mack et al. (2013) had relatively low coherence, where some stimuli were equally distant from the central tendency of their category and that of the other category. Bowman and Zeithamova (2018) used a similar model-based fMRI methodology, but a different category structure that included greater coherence of stimuli within each category. They found evidence consistent with the predictions of the prototype model in both behavior and the brain. Bowman and Zeithamova (2019) then tested the idea that concept learning may involve both specific (exemplar) and generalized (prototype) representations under different conditions and found that coherence of the underlying category structure is a critical factor determining the success of forming a generalized category representation (Minda and Smith, 2001).

Theoretically, the same network of regions could be involved in concept representation across different category structures, with only the nature of the representations formed being different. However, recent findings are more indicative of distinct neural mechanisms existing for representing specific instances versus summary representations, with these mechanisms being differentially engaged for different category structures. Mack et al. (2013) identified a network of regions representing specific exemplars that included lateral prefrontal and parietal regions implicated in supporting memory for specific events and preventing interference in episodic memory tasks (Badre and Wagner, 2005; Kuhl and Chun, 2014). Regions tracking prototype model predictions by Bowman and Zeithamova (2018) included hippocampus and ventromedial prefrontal cortex. Although no above-threshold evidence for exemplar representation during concept generalization was found in that study, subthreshold exemplar regions were consistent with those found by Mack et al. (2013) and distinct from those tracking prototype predictors. Thus, distinct neural mechanisms creating different types of may be engaged in concept learning, with their contribution varying across category structure. We will discuss further evidence for this notion from nonhuman primate research later in this review (Wutz et al., 2018).

The notion that both specific and generalized memories may represent concepts and support generalization aligns with findings from a different generalization task: episodic inference. In a typical task, participants learn a set of overlapping associations, such as A relates to B and B relates to C (Underwood, 1949). Although many behavioral and neuroimaging studies showed that reactivation of A during BC-learning can lead to interference and forgetting of C (Anderson, 2003; Kuhl et al., 2011), another outcome is episodic inference, where a relation is inferred between A and C. Although episodic inference can be achieved on-demand, from separate memories of individual events (Kumaran and McClelland, 2012), it can also result from memory integration, whereby new events are linked with prior related memories into a combined representation (Schlichting et al., 2014; Richter et al., 2016; Zeithamova and Preston, 2017). Notably, the same hippocampal-vmPFC interactions implicated in concept learning (Kumaran et al., 2009; Bowman and Zeithamova, 2018; Frank et al., 2019) have been implicated in memory integration and inference in both neuroimaging (Zeithamova et al., 2012; Schlichting et al., 2015) and lesion work (Dusek and Eichenbaum, 1997; Ryan et al., 2016; Spalding et al., 2018). Finally, the same regions have been shown to underlie schema-related memory (Tse et al., 2007, 2011; van Kesteren et al., 2012; Spalding et al., 2015; Brod et al., 2017; Gilboa and Marlatte, 2017; Gilboa and Moscovitch, 2017; Baldassano et al., 2018; Romero et al., 2019). Thus, hippocampal-vmPFC memory integration mechanisms may serve to link related information to a coherent representation in service of a range of generalization tasks, including concept learning.

Congruency and reactivation aid memory integration

Understanding how the brain links related information may help us to take a step further and focus on how memory integration processes can be enhanced, as desired, for example, in educational settings. Factors such as congruency between associates and reactivation of previously learned information have been shown to facilitate memory integration (Van Kesteren et al., 2018). However, how the brain achieves such memory improvements and how they link to memory integration processes is yet unknown.

To examine how enhanced memory integration is realized, van Kesteren et al. (2019) extended the AB-BC inference paradigm described in the previous section. They added a factor of congruency between A and C (a scene and an object), which is a strong enhancing factor in conventional associative memory experiments (van Kesteren et al., 2012, 2013a,b), and asked participants to rate the strength of reactivation (subjective reactivation of the scene) they experienced during BC-learning (pseudoword with object). Both these congruency and reactivation factors were expected to yield improved memory integration as was found in a recent experiment that used a more educationally valid version of this paradigm (van Kesteren et al., 2018).

This behavioral effect was replicated, and brain activity during memory integration was correlated with three behavioral factors: memory, congruency, and reactivation. MTL, including the hippocampus, was related to memory performance, whereas congruency was associated with activity in the vmPFC and the hippocampus, and reactivation strength revealed an extensive retrieval network that included the vmPFC and the hippocampus. Thus, hippocampus-vmPFC mechanisms may contribute to several aspects and modulatory factors involved in concept learning. This work extends prior work on schema-related memory benefits (Ghosh and Gilboa, 2014; Ryan et al., 2016) and provides new insights into how schema-congruency and active reactivation of existing knowledge improves memory and integration.

Integration of reward and concept representations in categorization

Although we so far discussed the role of vmPFC in memory integration, schema formation, and the acquisition of conceptual representations, a separate line of research on economic decision making emphasizes the role of vmPFC in reward processing (Kable and Glimcher, 2009). Reward is an important factor affecting categorization (Seger and Peterson, 2013), but has so far received little attention in neuroscientific studies of concept learning. Intriguingly, research into reward and value-based decision making has identified cortical regions that are also important for categorization (Summerfield and Tsetsos, 2012; Jocham et al., 2014). Here we focus on the vmPFC and an additional area, the intraparietal sulcus (IPS)/inferior parietal (IP), both of which are associated with both reward and concept representation. IPS/IP is active during categorization and is sensitive to category representation in humans (Seger et al., 2015; Wheeler et al., 2015) and nonhuman primates (Sarma et al., 2016).

Braunlich and Seger (2016; K. Braunlich & C.A. Seger, unpublished observations) examined how IPS/IP and vmPFC are involved in processing the sum of evidence for category membership and associated reward. Braunlich and Seger (2016) developed a task in which four features probabilistically related to category membership were presented in series over time. Bayesian model selection showed that the interaction between evidence and time to end of trial better accounted for activity than evidence alone in IPS/IP. K. Braunlich and C.A. Seger (unpublished observations) examined how reward availability affected evidence integration and decision during categorization using this task. Subjects were informed about when reward would be available; for example, correct responses would yield a reward when the third feature was presented, but not before or after. Activity in IPS/IP increased as a function of the amount of evidence and time to reward. In addition, the vmPFC and anterior hippocampus showed a ramping pattern of activity as time of reward approached, indicating representation of both the sum of evidence and the temporal distance for reward availability. Functional connectivity analysis found hippocampus connectivity with visual cortex, and vmPFC with both visual cortex and frontoparietal regions, indicating that the vmPFC may receive input from these regions that can be used to monitor the ongoing decision process and maximize reward.

Recent complementary studies have examined how reward expectation is integrated with representational knowledge during categorization. Braunlich et al. (2017) had subjects learn to categorize abstract polygonal stimuli formed as distortions of a prototypical stimulus according to two different decision criteria and found that both IPS/IP and vmPFC were sensitive to distance in perceptual space from the stimulus to the current criterion. C.A. Seger, K. Braunlich, and Z. Liu (unpublished observations) required subjects to combine categorization with information about reward probability to predict outcomes. On each trial, participants saw a cue indicating the probability that correct performance would be rewarded (0, 25, 50, 75, or 100%), then viewed and categorized a stimulus, and finally received reward. During stimulus presentation, IPS/IP was sensitive to the interaction of prototype distance and reward probability, and coded for reward, category, and their interaction in a representational similarity analysis. IPS/IP was also sensitive to both reward and prototype distance prediction error at the time of feedback, indicating a possible role in updating representations based on feedback. The vmPFC was sensitive to reward probability at cue and feedback, indicating a role in maintaining contextual information about reward probability across the trial and integrating it with categorical information. This work provides an intersection of vmPFC research on memory integration discussed here with otherwise separate research that emphasizes the role of vmPFC in subjective value representation through integration of multiple converging inputs (Clithero and Rangel, 2014; Berkman et al., 2017).

Modeling the role of the rostrolateral prefrontal cortex in category learning and generalization

Most concept-learning studies and general computational literature have focused on memory processes and similarity-based mechanisms: how memory representations are built around common information and retrieved on the basis of representational overlap to support novel judgments. Similarity-based processes are critical for long-term category representation (Nosofsky, 1986), however, many theories suggest that such processes are supplemented, in some cases, by inferential processes that are akin to using logical rules or reasoning strategies (Smith and Sloman, 1994; Ashby et al., 1998). For example, rules may augment categorization during early stages of learning while long-term representations are being formed and may be called upon again during generalization when people are confronted with stimuli that are ambiguous or similar to multiple previously acquired category representations (Nosofsky et al., 1994; Palmeri and Nosofsky, 1995; Erickson and Kruschke, 1998; Juslin et al., 2001).

How inferential categorization processes are instantiated in the brain is an open question. Broad neurobiological category-learning theories have focused on the lateral PFC as a whole in the operation of inferential and executive reasoning processes (Seger and Miller, 2010; Ashby and Maddox, 2011). However, they do not attempt to distinguish among subregions of the lateral PFC with respect to such processes. In the broader literature, abstract reasoning processes such as higher-level reasoning, problem solving, and analogy are often thought to depend on the rostrolateral PFC (rlPFC; Christoff et al., 2001; Kroger et al., 2002; Bunge et al., 2005; Green et al., 2006; Hampshire et al., 2011; Watson and Chatterjee, 2012). Evidence for this comes not only from activation-based studies of such abstract reasoning tasks, but also from anatomical data, such as the area's relative size in humans compared with other primates (Burgess et al., 2007) and its growth trajectory during development, which tracks the emergence of higher-level reasoning in children and adolescence (Dumontheil, 2014; Vendetti and Bunge, 2014).

Despite being associated with many abstract reasoning capabilities, there is yet to be an encompassing theory of rlPFC that explains its underlying computational contribution across domains. Hierarchical control theory has the most general account of rlPFC function, and suggests that it sits atop a rostrocaudal gradient of abstraction in the lateral cortex (Badre and D'Esposito, 2007, 2009). However, more recent work on control theory has challenged the notion that rlPFC is simply involved in abstract control processes, based on evidence that the rlPFC activates for a number of processes that are not easy to align with the idea that it is solely involved in abstract control (Badre and Nee, 2018). For example, it tends to be activated when people make the decision to explore new choice options rather than exploit options with the current highest expected value in reinforcement learning (Daw et al., 2006) and it has interesting temporal trajectories across sequential tasks that do not vary temporally in their control demands per se (Desrochers et al., 2015, 2019).

Recent research in category learning may shed light on how the rlPFC diverges from the other, more caudal, areas of the lateral PFC. Converging evidence across a number of studies suggests that the rlPFC instantiates an inferential process that is sensitive to the novelty and decisional uncertainty associated with a stimulus. Specifically, people will tend to engage rules when confronted with a stimulus that is both novel and difficult to categorize given the previously learned category representations. In support of this hypothesis, using an iterative rule-learning task, Paniukov and Davis (2018) found that the rlPFC is engaged early in learning and remains engaged as long as uncertainty remains about the correct category rule. Davis et al. (2017) showed that the rlPFC was more engaged during acquisition of relational category-learning rules than acquisition of feature-based category rules. Later, once the relational rules were well learned, the rlPFC was not more activated for relational rules across the board, but instead was more activated to the extent that stimuli were novel examples of the previously learned relations. Finally, O'Bryan et al. (2018) showed that the rlPFC tracked the contribution of dissimilarity-based heuristics during generalization to novel ambiguous items. Such dissimilarity-based processes are thought to be based on higher-level inferential strategies. Together, these results suggest that the rlPFC integrates decisional information and stimulus novelty to determine when to use inferential processes in category learning and generalization.

Different prefrontal cortex dynamics for learning at different levels of abstraction

Many studies support the view that the lateral PFC plays a central role for concepts when they require high levels of abstraction (Badre and D'Esposito, 2009) and rule-based reasoning (Davis et al., 2017; O'Bryan et al., 2018). But lateral PFC has also been implicated in low-level, similarity-based categorization (Davis et al., 2017; O'Bryan et al., 2018) and representation of specific exemplars (Mack et al., 2013). Bowman and Zeithamova (2018, 2019) suggested that distinct processes and types of representations can be involved in concept learning, with their relative contribution depending on the coherence among category members. Lower coherence (lower similarity among category members) makes the formation of generalized representations based on similarity difficult and may require a higher level of abstraction. However, are high-level abstractions simply “more of the same”, using the same mechanisms and networks as low-level abstractions? Or do low- and high-level abstractions involve distinct anatomical circuits and functional mechanisms in the PFC?

Important insights for resolving this question can come from neurophysiological recordings, which provide single-neuron resolution, but can also provide information about how groups of neurons communicate through analysis of their collective activity in local field potentials. To investigate how category abstraction is organized in prefrontal cortex, Wutz et al. (2018) recorded from multielectrode arrays in lateral PFC and trained monkeys in a dot-pattern category task (Posner and Keele, 1968) that provides a parametrically controlled and mathematically straightforward way to separate categories by level of abstraction. On each recording day, the monkeys learned to categorize dot patterns (i.e., exemplars), which were created by spatially distorting an underlying prototype pattern. Two new, pseudorandomly generated category prototypes were used per recording day, each of which was organized in a sequence of training blocks containing an increasing number of exemplars. By this, Wutz et al. (2018) made sure that the monkeys generalized over a large pool of exemplars (i.e., 64–256 per category) and eventually learned to extract the underlying category prototype. Critically, varying the degree of spatial distortion of the exemplars from the prototype allowed the authors to control the required level of abstraction for the category decision. Low-distortion exemplars look alike and can be categorized based on the similarity of their sensory features. High-distortion exemplars, however, can look very different from each other requiring greater abstraction.

Wutz et al. (2018) found category abstraction organized in different subregions in the PFC. The ventrolateral PFC (vlPFC) was more engaged for low-level abstractions, whereas the dorsolateral PFC (dlPFC) processed more high-level abstractions. Beyond this anatomical distinction, however, the combined analyses of spiking activity and local field potentials also suggested distinct functional mechanisms based on different temporal dynamics and frequency characteristics. In terms of frequency, there is growing evidence that gamma versus beta oscillations are involved in bottom-up versus top-down processing, respectively (Buschman and Miller, 2007; Jensen et al., 2007; Engel and Fries, 2010). Category signals for low-level abstractions in the vlPFC were found in gamma oscillations (60–160 Hz), evoked potentials, and spiking activity when the exemplars were shown and their sensory properties were processed. In contrast, the dlPFC showed stronger category signals for high-level abstractions in beta oscillations (10–35 Hz) throughout the memory delay epoch when the exemplars had to be kept in mind. Moreover, spiking activity in the dlPFC during the delay occurred at specific phases of its ongoing beta-band oscillations. This pattern suggested that neurons in the vlPFC were more driven by bottom-up inputs, whereas in the dlPFC neurons fired more in sync with its inherent, top-down dynamics during category processing.

The study demonstrated that distinct neural circuits (vlPFC vs dlPFC) communicate through distinct frequency channels (gamma vs beta) and at different times (sample vs delay epoch) when inferring regularities about the world on low versus high levels of abstraction (Fig. 2). By extension, these findings support the existence of two distinct functional mechanisms (bottom-up vs top-down) for category abstraction in the PFC. Gamma oscillations are typically linked to the feedforward flow of cortical activation (Bastos et al., 2015; Jensen et al., 2015) and the vlPFC receives direct inputs from inferior temporal cortex (O'Reilly, 2010), potentially continuing its functional properties into prefrontal cortex function. Therefore, category processing through the vlPFC-gamma network may be viewed as an object-recognition/pattern-matching problem governed by bottom-up principles and subserving low-level abstraction. By contrast, beta oscillations support cortical feedback (Bastos et al., 2015; Jensen et al., 2015) and there is stronger connectivity between the dlPFC and parietal cortex (O'Reilly, 2010). Thus, the dlPFC-beta network may implement top-down, experience-based generalization and identify more abstract, conceptual relationships that transcend appearance beyond object recognition. Given that different regions of the lateral PFC contribute to category formation through distinct functional circuits because of their differential connectivity to posterior cortex, a critical goal for future research will be to delineate whether the lateral PFC is organized on the basis of representation type (rule vs similarity-based) or if more general control mechanisms define its topography (e.g., gating or branching; Badre and Nee, 2018).

• Download figure
• Open in new tab
• Download powerpoint

Figure 2.

A schematic depiction of results from Wutz et al. (2018). Distinct neural circuits communicate through distinct channels when concept learning requires low levels of abstraction versus high levels of abstraction.

In conclusion, the current review highlights representative samples from a recent boom of concept learning studies. By leveraging well developed computational models to interrogate neural mechanisms and representations, this work has implicated a broad network of brain regions including the hippocampus, PFC, and parietal cortices. Importantly, this work has significantly advanced our understanding of concept learning by characterizing the nature of the component mechanisms and their underlying neural machinery. The result is a converging neurocomputational account of concept learning that integrates brain systems involved in attention, memory, reasoning, cognitive control, and reward processing. Theoretically, this work brings resolution to the decades-long debate on the nature of category representations and emphasizes the need for comprehensive theories that bridge brain and behavior. Looking forward, understanding how these multiple brain systems interact throughout learning to support the flexible formation and use of concept knowledge is a key research aim. Notably, research in concept learning is well positioned to reach the holy grail of the computational model-based neuroscientific approach: not only are formal theories used to inform understanding of neural mechanisms, findings from the brain will undoubtedly motivate meaningful extensions to formal theories.