Categorization involves grouping objects together according to perceptual or relational similarity. This requires mechanisms that can simultaneously generalize across within-category differences (e.g., different dog breeds vary in head shape, body size, and fur) and discriminate against between-category similarities (e.g., dogs and cats have similar body structure). Balancing generalization and discrimination can be accomplished by the hippocampus, which has been shown to 1) link experiences together according to overlapping features and 2) amplify differences between relatively similar memory traces (McNaughton and Morris, 1987, O'Reilly and McClelland, 1994; Hunsaker, 2013).

Early theories of categorization minimized the importance of the hippocampus in category learning (Ashby, Alfonso-Reese, Turken, & Waldron, 1998). This was largely because patients with amnesia did not show reliable learning impairments across multiple categorization tasks (Knowlton and Squire, 1993, Knowlton et al., 1996, Filoteo et al., 2001, Haslam, 1997; but see Zaki, 2004). However, more recent evidence from neuroimaging (Zeithamova et al., 2012, Kumaran et al., 2009, Mack et al., 2016), neurophysiology (Hampson et al., 2004, Kraskov et al., 2007, Kreiman et al., 2000), and rodent inactivation studies (Kim, Castro, Wasserman, & Freeman, 2018) have challenged this idea and argue that the hippocampus is central to categorization. Now, it is predicted that the hippocampus builds and maintains flexible category representations (Mack et al., 2018, Bowman and Zeithamova, 2018). This function mirrors the role of the hippocampus in maintaining structured memory representations, called ‘schemas’ (Tse et al., 2007, Baraduc et al., 2019, Guo et al., 2023).

This new view has led to the development of theoretical frameworks that describe how well-documented mechanisms of the hippocampus could be leveraged during category learning. For example, EpCon (Episodes-to-Concepts), describes how pattern separation (i.e., separating similar memory traces to avoid interference; Marr, 1969, Leutgeb et al., 2007, Bakker et al., 2008, Yassa and Stark, 2011, Kirwan et al., 2012), pattern completion (i.e., using partial information to retrieve memory traces; Horner et al., 2015, Gold and Kesner, 2005, Guzman et al., 2016), and memory integration (i.e., integrating new memory traces into existing representations; Dusek and Eichenbaum, 1997, Eichenbaum, 2001, Backus et al., 2016, Schlichting and Preston, 2015, Pajkert et al., 2017) could all be relevant for learning new categories (Mack et al., 2018). EpCon posits that the hippocampus 1) retrieves memory representations that are similar to the stimulus being categorized (i.e., pattern completion), 2) integrates new stimuli into existing representations (i.e., memory integration), and 3) forms new representations after encountering surprising stimuli (i.e., pattern separation). Frameworks like EpCon are intuitive in that they build on decades of research. Nevertheless, few experiments have tested these predictions directly.

One approach to test the EpCon framework is to utilize a computational model of categorization that encompasses fundamental mechanisms of the hippocampus. One such model is SUSTAIN (Fig. 1; Supervised and Unsupervised STratified Adaptive Incremental Network; Love et al., 2004, Love and Gureckis, 2007). SUSTAIN assumes that similar training experiences tend to cluster together in memory (Fig. 1A). Categories are represented by single or multiple ‘clusters’, where each cluster reflects a learned group of similar training experiences (Fig. 1B). Categorizing a new stimulus involves retrieving cluster representations that are perceptually similar to that stimulus (i.e., pattern completion; Fig. 1C). After receiving feedback, the cluster representations are updated by 1) integrating the new stimulus into existing clusters (i.e., memory integration; Fig. 1D) and/or 1) forming a new cluster (i.e., pattern separation; Fig. 1E). We posit that SUSTAIN is a desirable model to bridge the fundamental mechanisms of the hippocampus with principles of category learning.

Indeed, there is growing evidence that activity in the hippocampus is functionally similar to the clustering mechanism of SUSTAIN. Multiple studies have demonstrated that the hippocampus creates ‘cognitive maps’ (Tolman, 1948, Behrens et al., 2018) of non-spatial, multidimensional feature spaces (Eichenbaum and Cohen, 2014, Theves et al., 2019, Solomon et al., 2019, Constantinescu et al., 2016, Morton et al., 2017). These representations emphasize category-relevant stimulus information and reflect task goals (Theves et al., 2020, Mack et al., 2016). Furthermore, Mok & Love, 2019 showed that a clustering model could simulate neural activity of place cells and grid cells as a rat navigated an environment. This suggests that similar mechanisms may be recruited to mediate both spatial navigation and concept learning. Expanding the investigation of the hippocampus to non-spatial paradigms like categorization may provide key insight regarding generalized hippocampal mechanisms that go beyond spatial navigation.

In the current experiment, we used inhibitory DREADDs (Designer Receptors Exclusively Activated by Designer Drugs; Roth, 2016) to examine the role of the dorsal hippocampus (HPC) in category learning. Using a touchscreen apparatus, rats were trained to categorize distributions of controlled visual stimuli derived from classic human paradigms that have been used for decades (Ashby et al., 1998). The category stimuli contained black and white gratings that varied along two continuous dimensions (i.e., spatial frequency and orientation; Fig. 2A; Broschard et al., 2019, Ashby et al., 1998). For some rats, categorizing the stimuli encouraged a shift of attention to a single stimulus dimension (i.e., 1D tasks; spatial frequency or orientation). For other rats, categorizing the stimuli required attention to both stimulus dimensions (i.e., 2D tasks; spatial frequency and orientation). Inactivation of the HPC impaired category learning and generalization for both the 1D tasks and 2D tasks. We then fit SUSTAIN to the learning data to test the role of the HPC in storing and retrieving category representations.