The neural mechanisms underlying categorization are complex and likely involve multiple neural substrates, among these include the basal ganglia (Seger and Miller, 2010, Zeithamova et al., 2019, Ashby et al., 1998; Ashby & O’Brien, 2005). In a seminal paper, Knowlton, Mangels, & Squire (1996) had patients with Parkinson’s Disease learn the Weather Prediction task, a probabilistic categorization task in which participants learn to combine information from sets of cards to predict the weather (i.e., ‘sunny’ or ‘rainy’). Compared to healthy comparisons, the patients with Parkinson’s Disease were impaired to learn this task, suggesting that the basal ganglia are necessary for categorization. Since this seminal work, the importance of the basal ganglia has been replicated by neuroimaging (Seger & Cincotta, 2002), neuropsychological (Maddox, Aparicio, Marchant, & Ivry, 2005), and neurophysiological experiments (Antzouloatos & Miller, 2014).

Multiple functions have been proposed regarding the role of the basal ganglia in categorization, including selective attention (Swainson et al., 2006, Brown and Marsden, 1988), set shifting (Owen et al., 1993, Volz et al., 1998, Lombardi et al., 1999), feedback processing (Schultz et al., 1997, Shohamy et al., 2008), and stimulus generalization (Seger & Cincotta, 2005; see Shohamy et al., 2008, Ashby and Ennis, 2006 for reviews). The diversity of these functions has been explained by assuming that each function is mediated through a distinct corticostriatal loop (Seger, 2008, Yahya, 2020). For instance, the head of the caudate nucleus may be critical for executive functioning via connections to the prefrontal cortex (i.e., the executive loop), whereas the tail and body of the caudate nucleus may be critical for stimulus generalization via connections to the visual cortex (i.e., the visual loop; Seger and Cincotta, 2005, Lopez-Paniagua and Seger, 2011). Under this framework, the executive loop uses selective attention and feedback processing to test simple category rules (see Antzoulatos and Miller, 2011, Antzoulatos and Miller, 2014, Villagrasa et al., 2018, Shohamy et al., 2004, Schultz et al., 1997, Shohamy et al., 2008). The visual loop, on the other hand, relies on the large convergence of sensory information from the visual cortex, allowing for generalization across stimuli that are perceptually similar (Seger, 2013).

These hypotheses have been formalized by a model of human category learning, COVIS (COmpetition between Verbal and Implicit Systems; Ashby et al., 1998). COVIS posits that humans learn new categories through two independent systems: the declarative system and the procedural system. The declarative system utilizes the executive loop and tests simple category rules, whereas the procedural system utilizes the visual loop and learns to incrementally associate category stimuli to appropriate behavioral responses. For decades, the predictions of the COVIS model have been tested by training participants to categorize distributions of visual stimuli that change along two continuous dimensions (e.g., spatial frequency and orientation; Figs. 1A-B; Ashby & Maddox, 2005). For some tasks, the categories can be learned using a unidimensional rule (i.e., rule-based; RB tasks; Fig. 1A). For other tasks, the categories must be learned by combining information from both stimulus dimensions (i.e., information integration; II tasks). Importantly, COVIS makes specific predictions such that the head of the caudate nucleus (and the declarative system) is critical to learn the RB tasks, whereas the tail of the caudate nucleus (and the procedural system) is critical to the learn the II tasks.

Support for these predictions is mixed. Patients with Parkinson’s Disease are typically impaired to learn RB tasks (Maddox, Aparicio, Marchant, & Ivry, 2005; see Price, 2006, Ashby et al., 2003 for reviews) and not II tasks (Ashby et al., 2003, Filoteo et al., 2005). Impairments in the RB tasks are generally attributed to deficits in selective attention (Filoteo, Maddox, Ing, Zizak, & Song, 2005, 2007; Filoteo, Maddox, & Davis, 2001). A few neuroimaging studies have found dissociable basal ganglia activity in the RB and II tasks that aligns with the predictions of COVIS (Soto, Waldschmidt, Helie, & Ashby, 2013; Nomura et al., 2007); however, other studies do not replicate these results (Carpenter et al., 2016, Milton and Pothos, 2011). In sum, there is some support for the COVIS model; however, a more thorough test of these predictions may be necessary.

To clarify the roles of the basal ganglia in RB and II learning, we lesioned subregions of the rodent striatum. Multiple anatomical studies have compared homologies between the primate and rodent striatum (e.g., Heilbronner et al., 2016, Balsters et al., 2020). The head of the caudate nucleus in primates is likely homologous to medial portions of the rodent striatum (i.e., dorsomedial striatum; DMS), characterized by direct inputs from the prefrontal cortex. The tail of the caudate nucleus in primates is likely homologous to lateral portions of the rodent striatum (i.e., dorsolateral striatum; DLS), characterized by sensory and sensorimotor inputs (Folsters et al., 2021; West et al., 1990). These homologies are supported by work comparing the functional similarity between the human and rodent striatum in the context of action selection (Balleine & O’Dohery, 2009; Yin & Knowlton, 2006). Specifically, the DMS and head of the caudate nucleus are both critical for selecting goal-directed behaviors, whereas the DLS and tail of the caudate nucleus are both critical for mediating habitual behaviors.

In the current experiment, we trained rats to learn adapted versions of the RB and II tasks using a touchscreen apparatus (i.e., 1D tasks and 2D tasks, respectively; Broschard et al., 2019, Broschard et al., 2020, 2021) to examine the roles of the DMS and DLS in category learning. Computational modeling simulated the functions of the DMS and DLS during category learning (Love et al., 2004, Broschard et al., 2021, Broschard et al., 2019). We found that rats with lesions to the DMS were impaired on both task types, whereas lesioning the DLS had no effect on learning. Model simulations suggest that the DMS serves a general role in category learning by mapping category representations to appropriate behavioral responses. These results do not support the predictions of COVIS.