A neural network model of hippocampal contributions to category learning

Article activity feed

1. 

   Version published to 10.7554/elife.77185 on eLife

   Dec 11, 2023
2. 

   eLife

   Aug 10, 2023

   Author Response

   Reviewer #1 (Public Review):

   Sučević and Schapiro investigated a neurobiologically inspired model of human hippocampal structure and computation in category learning. In three separate simulations, the model (CHORSE) is presented with learning tasks defined by various category structures from prior work and evaluated for its ability to learn the category structure, generalize categorization to novel stimuli, and accurately recognize previously encountered stimuli. Although originally conceived of as a computational model of associative memory, C-HORSE is demonstrated to quite naturally account for human-like learning of the three category tasks. Notably, the authors characterize the mechanisms underlying the model's learning by way of additional simulations in which "lesions" to the model's monosynaptic pathway (MSP; …

   Author Response

   Reviewer #1 (Public Review):

   Sučević and Schapiro investigated a neurobiologically inspired model of human hippocampal structure and computation in category learning. In three separate simulations, the model (CHORSE) is presented with learning tasks defined by various category structures from prior work and evaluated for its ability to learn the category structure, generalize categorization to novel stimuli, and accurately recognize previously encountered stimuli. Although originally conceived of as a computational model of associative memory, C-HORSE is demonstrated to quite naturally account for human-like learning of the three category tasks. Notably, the authors characterize the mechanisms underlying the model's learning by way of additional simulations in which "lesions" to the model's monosynaptic pathway (MSP; direct connections between ERC and CA1) are contrasted with lesions to its trisynaptic pathway (TSP; pathway connecting ERCDG-CA3-CA1). These in silico lesions offer key insight into the computational principles underlying theorized hippocampal functions in category learning: whereas MSP provides incremental learning of shared features diagnostic to category membership that are important for category generalization, TSP learns item-specific information that drives recognition behaviour. The authors propose that C-HORSE's successful account of a broad set of category learning datasets provides clear support for the role of complementary hippocampal functions mediated by MSP and TSP in category learning. This work adds compelling computational evidence to a growing literature linking hippocampus to a broader role in cognition that extends beyond declarative memory.

   The model simulations are clear and properly conducted. The three datasets examined offer a relatively broad set of findings from the category learning literature; that the models provide reasonable accounts of human performance in all three speaks to the model's generalizability. Overall, I find this work exciting and an important step in linking longstanding well-established formal learning theories of psychology with neurobiological mechanism. Several weaknesses dampen this excitement, each of which are detailed below:

   1. C-HORSE is presented as a new entry into a rich field of formal computational models of category learning. As noted above, the datasets examined span a broad range of learning contexts and structures and the model's ability to account for learning behaviour is compelling. However, no other models are leveraged to perform a direct evaluation. In other words, CHORSE's predictions are compelling, but is it better than other competing models in the literature? To be clear, C-HORSE offers a novel alternative with its fundamental mechanisms originating from anatomical structure and connectivity. As such, a proof-of-concept showing that such a neurobiologically inspired framework can account for category learning behaviour is a worthwhile contribution in its own right and a clear strength of this paper. However, how to consider this model relative to existing theoretical frameworks is not well described in the manuscript.

   We very much appreciate this point — see response to Editor summary point #3 above.

   1. Relatedly, C-HORSE is evaluated in terms of qualitative fit to behaviour measures from prior studies and in all three simulations restricted to measure of end of learning performance. Again, an appeal to the proof-of-concept nature of the current work may provide an appropriate context for this paper. But, a hallmark of well-established category learning models (e.g., SUSTAIN, DIVA, EBRW, SEA, etc.) is their ability to account for both end of learning generalization (and in some cases, recognition) and behaviour throughout the learning process. C-HORSE does provide predictions of how learning unfolds over time, but how well this compares to human measures is not considered in the current manuscript. Such comparisons would strengthen the support for C-HORSE as a viable model of category learning and help position it in the busy field of related formal models.

   We completely agree about the value of this, and we have added empirical timecourse data for comparison with all simulations, as described in response to Editor summary point #7, above.

   1. A consistent finding across all three simulations is that the TSP provides item-specific encoding. Evidence for this can be inferred by contrasting categorization and recognition performance across the TSP- and MSP-only model variants. In the discussion, the authors draw a parallel between exemplar theories of category learning and the TSP, which is a compelling theoretical position. However, as noted by the authors, unlike exemplar theories, the TSP-only model was notably impaired at categorization. The author's suggestions for extensions to CHORSE that would enable better TSP-based categorization are interesting. But, I think it would be helpful to understand something about the nature of the representations being formed in the TSP-only model. For example, are they truly item-specific, are the shared category features simply lost to heightened encoding of item-unique features, are category members organized similarly to the intact model just with more variability, and so on. Characterizing the nature of these representations to understand the limitations of the TSP-only model seems important to understanding the representational dynamics of C-HORSE, but are not included in the current manuscript.

   The RSA results, now included for Simulations 2 and 3 in addition to Simulation 1, provide the information needed to characterize the nature of the TSP representations. Generally speaking, they are truly item specific, meaning that each item is represented by its own distinct set of units. This is a demonstration of the classic pattern separation function of this pathway, taking similar inputs and projecting them to orthogonal populations of neurons. Simulation 1 is the clearest example of this, where there is virtually no similarity and very low variability in the item similarity structure in DG and CA3. The new Simulation 3 RSA shows us where the limit is to this pattern separation ability of the TSP, with highly typical items being represented by somewhat overlapping populations of neurons in DG and CA3. To the extent that the TSP can succeed in generalization, it seems to involve this pattern separation failure.

   We have made these points more explicit in new discussion of the RSA results:

   • Simulation 1: “In the initial response, there was no sensitivity at all to category structure in DG and CA3 — items were represented with distinct sets of units. This is a demonstration of the classic pattern separation function of the TSP, applied to this domain of category learning, where it is able to take overlapping inputs and project them to separate populations of units in DG and CA3.” • Simulation 3: “As in the prior simulations, DG and CA3 represented the items more distinctly than CA1, and settled activity after big-loop recurrence increased similarity, especially in CA1. This simulation was unique, however, in that DG and CA3 showed clear similarity structure for the prototype and highly prototypical items. There is a limit to the pattern separation abilities of the TSP, and these highly similar items exceeded that limit. This explains why, at high typicality levels, the TSP could be quite successful on its own in generalization (Figure 5e), and why it struggled with atypical feature recognition for these items (Figure 5f).”

   1. In general, a detailed description that links model mechanisms and analyses to the learning constructs of interest for the different simulations is lacking. For example, RSA results for simulation 1 are contrasted for initial and settled representations, but what is meaningful about these two timepoints is not directly stated (moreover, what initial and settled response mean in terms of the current model is not explained). The authors do briefly suggest that differences between initial and settled representations may reflect encoding dynamics before and after bigloop recurrence, but this is not established as a key metric for evaluating the nature of the model representations. In general, more motivation is needed to understand what the chosen analyses reveal about the nature of the model's learning process and representations.

   We have added more description of the motivation for our analyses. See response to Editor summary point #6 above.

   1. I appreciate the comparison in the discussion to extant models of categorization. Certainly, the exemplar and prototype models are fixtures of the category learning literature and they somewhat align with the type of learning that TSP and MSP, respectively, provide. REMERGE and SUSTAIN are also briefly mentioned, but their discussion is limited which is unfortunate as they are actually more functionally equivalent to C-HORSE. I think, however, that the authors are missing an opportunity to discuss how C-HORSE offers a means for bridging levels of analysis to connect neurobiological mechanisms with these notably successful psychological models of category learning. Rather than framing C-HORSE as a competitor to existing models, it should be viewed as an account existing on a different level of analysis. In this sense, it complements existing approaches and potentially extends a theoretical olive branch between the psychology and neuroscience of category learning.

   We love this point about bridging levels of analysis and have added it to our discussion of the model’s relationship to other models, see Editor summary point #3 above.

   1. The discussion takes a broad perspective on covering evidence concerning hippocampal contributions to category learning. Although comprehensive, some sections are not well connected back to the main thrust of the paper. For example, a section on neuropsychological accounts of the hippocampus and category learning summarizes central aspects of this literature but is never reflected on through the lens of the current findings. I do think this prior work is relevant, especially since it a central theme of the hippocampus not being necessary for category/concept learning, but its connection back to the current study is not well argued. Similarly, the section on consolidation and sleep is relevant, but in its current form does not seem to fit with the rest of the paper.

   We have implemented these suggestions through very significant revisions to the Discussion. We now better connect the sections to the main argument of the paper and made cuts throughout, including removing the section on consolidation and sleep.

   Reviewer #2 (Public Review):

   The authors present a model of the hippocampal region that incorporates both the (indirect) trisynaptic and (direct) mono-synaptic pathways from entorhinal cortex (EC) to CA1 - the former incorporating projections from EC to dentate gyrus (DG), DG to CA3, and CA3 to CA1, and exhibiting a higher learning rate. They demonstrate that exposing this network to stimuli consistent with standard empirical tests of category learning (e.g. where within-category exemplars share a set of common features) allows the network to reliably assign both novel and previously encountered stimuli to the correct category (e.g. the network can learn to classify stimuli and generalise this knowledge to new examples). They show that the tri-synaptic pathway (TSP) preferentially supports the encoding of individual exemplars (e.g. analogous to episodic memory) while the mono-synaptic pathway (MSP) preferentially supports category learning.

   The manuscript is well written, the simulation details appear sound, and the results are clearly and accurately presented. This model builds on a long tradition of computational modelling of hippocampal contributions to human memory function, strongly grounded in anatomical and electrophysiology data from both rodents and humans, and is therefore able to link phenomena at the level of individual cells and circuits to emergent behaviour - a major strength of this, and similar, work. However, I have two major concerns relating to the relationship between these findings and previously published work by the same and other authors.

   First, it is not clear to me - from the manuscript - whether these results represent a significant novel advance on previous publications from the same senior author. Figures 1 and 3D are almost identical to figures published in Schapiro et al. (2017) Phil Trans B, and the take-home message (that the MSP might support statistical learning) is the same. In brief, it seems that the authors have subjected an identical network to some new (but related) tasks and reached the same set of conclusions. I see no distinction between learning to extract 'statistical regularities' (in previous work) and learning 'the structure of new categories' (described here). As an aside, demonstrating that an autoencoder network can learn stimulus categories and generalise to new exemplars is also well established.

   We appreciate the opportunity to better articulate the novelty and importance of applying the model to the domain of category learning. There are crucial differences between statistical learning and category learning that make these simulations nontrivial (it did not have to be the case that the results would replicate for these category learning paradigms), and, importantly, many of the insights in the current work are category-learning specific (e.g., the effects of atypical features, trade-offs between generalization and recognition of exemplar-specific features). On the other hand, we of course agree that there are principles in common between statistical learning and category learning that are leading to the consistent findings. We added new material to the Introduction to explain the importance of these new simulations in the domain of category learning, and the value we see in demonstrating convergence across domains. See response to Editor point #1 above.

   Second, I have some concerns with the relationship between the properties of this hippocampal network model and well described properties of single cells in the rodent and human hippocampus. In particular, the CA1 units in this model (and to some extent, also the CA3 units) come to respond strongly to all exemplars from within each category (e.g. as shown in Figure 3D, bottom right panel). This appears to be at odds with the known properties of place and concept cells from the rodent and human hippocampus, respectively, which show little generalisation across related concepts (i.e. the Jennifer Aniston neuron does not fire in response to other actors from Friends, for example). If the emergent properties of this model are not consistent with existing data, then it is not a valid model.

   We appreciate the opportunity to discuss connections to the physiology literature. See response to Editor summary point #2 above.

   More generally, the authors are clear that this model is "a microcosm of [the] hippocampusneocortex relationship" and that the properties of the MSP "mirror those of neocortex". Why not assume that category learning is supported by an interaction between hippocampus and neocortex, then, as in the complementary learning systems (CLS) model? Aside from some correlational fMRI data and partial deficits in hippocampal amnesics - either of which could have a myriad of different explanations - what empirical data is better accounted for by this model than CLS? Put differently, what grounds are there for rejecting the CLS model? To some extent, this model appears to account for less empirical data than CLS, with the exception of a few recent neuroimaging studies (which are hard to interpret at the level of single cells)

   This is an important point for us to clarify, so we very much appreciate this comment. The crucial issue with CLS that motivated the microcosm theory is that the neocortex in the CLS framework learns far too slowly to support the kind of category learning studied in these paradigms, which unfolds over the course of minutes or hours. The neocortex in CLS was proposed to learn novel structure across days, months, and years.

   We have added the following to the Introduction:

   • “Despite its analogous properties, the MSP is not redundant with neocortex in this framework: the MSP allows rapid structure learning, on the timescale of minutes to hours, whereas the neocortex learns more slowly, across days, months, and years. The learning rate in the MSP is intermediate between the TSP (which operates as rapidly as one shot) and neocortex. The proposal is thus that the MSP is crucial to the extent that structure must be learned rapidly.”

   We also have this description in the Discussion:

   • “The MSP in our model has properties similar to the neocortex in that framework, with relatively more overlapping representations and a relatively slower learning rate, allowing it to behave as a miniature semantic memory system. The TSP and MSP in our model are thus a microcosm of the broader Complementary Learning Systems dynamic, with the MSP playing the role of a rapid learner of novel semantics, relative to the slower learning of neocortex.”

   Reviewer #3 (Public Review):

   The current work aimed to determine how the hippocampus may be able to detect regularities across experiences and how such a mechanism may serve to support category learning and generalization. Rapid learning in the hippocampus is critical for episodic memory and encoding of individual episodes. However, the rapid binding of arbitrary associations and one-shot learning was long thought suboptimal for finding regularities across experiences to support generalization, which were instead ascribed to other, slower-learning memory systems. More recent work has started to highlight hippocampal role in generalization, renewing the question of how generalization can be accomplished alongside memory for episodic details within a single memory structure. The current paper offers a reconciliation, presenting a biologically-inspired model of the hippocampus that is able to learn categories alongside stimulus-specific information comparably to human performance. The results convincingly demonstrate how distinct pathways within the hippocampus may differentially serve these complementary memory functions, enabling the single structure to support both episodic memory and categorization.

   Major strengths and contributions

   The paper includes simulation of three distinct categorization tasks, with a clear explanation of the unique aspects of each task. The key results are consistent across tasks, lending further support to the main conclusions of the role of distinct hippocampal pathways in learning specific details vs. regularities. Together with prior work on how the same architecture can support statistical learning in other types of tasks, this work provides important evidence of the broad role of the hippocampus in rapid integration of related information to serve many forms of cognition.

   Throughout the paper, the authors nicely explain in conceptual terms how the same underlying computations may serve all three categorization tasks as well as statistical learning and episodic inference tasks. Thus, the paper will be of broad interest, beyond researchers focused on modeling and/or categorization.

   On a conceptual level, this work provides a fruitful framework for understanding hippocampal functions, representations and computations. It provides a highly plausible mechanistic explanation of how category learning and generalization can be accomplished in the hippocampus and how distinct types of representations may emerge in distinct hippocampal subfields. The framework can be used to derive new testable predictions, some of which the authors themselves introduced. It also provides new insights into how the outputs of different pathways influence each other, providing a more nuanced view of the division of labor and interactions between hippocampal subfields. For example, the big loop recurrence would eventually lead to category influences even on the initially sparse, pattern separated representations in the CA3, which is an idea consistent with empirical observations.

   The presented computational model of the hippocampus is currently the most detailed and biologically plausible hippocampal model easily applicable in the area of cognitive neuroscience and behavioral simulations. The commonalities and differences with other related models (conceptual and computational) are well explained. Both the conceptual and technical descriptions of the model are exceptionally clear and detailed. The model is also publicly available for download for any researcher to use with their own task and data. All these aspects make it likely that other researchers may adopt the model in a wider range of tasks, stimulating new discoveries.

   The autoencoder nature of the model and the use of categorization tasks meant that some measures of interest, like recognition of exemplar-specific information, could not be evaluated by direct reading of the output layer to compare with some label (like old/new). The authors however came up with clever ways how to evaluate recognition performance in each task that was sensible and highlighted the multiple ways how one may think about information contained in neural representations in each layer. This approach can also be utilized by others for evaluating item-specific and category information in activation patterns, for example in analyses of fMRI.

   Finally, I thought the current paper and provided model may also serve as an excellent introduction to computational modeling for those new to this approach. The exceptional clarity of the conceptual and technical description of this model and the clear logic of how one may model a cognitive task and interpret results made this paper fairly accessible. Furthermore, the paper offered new insights and predictions based on analyzing the model's hidden layers, lesion performance, and/or noting some patterns of behavior unique to specific tasks. This was also instructive for highlighting the distinctive contributions that the computational modeling approach can have for furthering our understanding of cognition and the brain.

   We are extremely appreciative of the value the Reviewer sees in this work.

   Weaknesses

   The paper's strengths far outnumbered the weaknesses, that are minor. For one, the selected categorization tasks nicely complemented each other, but only covered stimuli with discretevalue dimensions (features like color, shape, symbol, etc). The degree to which the results generalize (or not) to continuous-value stimuli and different category structures (for instance information-integration or rule-based in COVIS framework) is not clear. How the model could be adjusted for continuous-value stimuli was not specified.

   We agree that the simulation of only discrete valued dimensions is a limitation. We chose to do this simply because it is easier to use discrete values in the model as currently implemented, but future work will certainly need to test whether the model can simulate the various paradigms that make use of continuous-valued dimensions. We have added an explicit acknowledgement of this issue in the Methods:

   • “The inhibition simulates the action of inhibitory interneurons and is implemented using a set-point inhibitory current with k-winner-take-all dynamics (O’Reilly, Munakata, Frank, Hazy, & Contributors, 2014). All simulations involved tasks with discrete-valued dimensions, as these are more easily amenable to implementation across input/output units whose activity tends to become binarized as a result of these inhibition dynamics. It will be important for future work to extend to implementations of category learning tasks with continuous-valued dimensions.”

   There is compelling evidence for the dissociation between different hippocampal pathways and subfields (CA1 vs. CA3) that the model is based on. As the authors noted, there is also compelling evidence for functional dissociations along the long hippocampal axis, with anterior portions more geared towards coarse, generalized representations while posterior towards more detailed, specific representations. The authors nicely pointed out that these proposals of withinhippocampus division of labor are less orthogonal than they may first appear, as there is greater proportion of CA1 in the anterior hippocampus. However, it is premature to imply that this resolves the CA1/CA3 vs. anterior/posterior question; the idea that existing anterior findings may be simply CA1 findings is currently only speculation. Furthermore, first studies indicating that anterior/posterior representational gradients may exist within each subfield are beginning to emerge.

   We completely agree that this is speculative at this point, which needed acknowledgment. See response to Editor summary point #2 above.

   Read the original source
3. 

   Version published to 10.1101/2022.01.12.476051v2 on bioRxiv

   Jul 24, 2023
4. 

   eLife

   Apr 26, 2022

   Evaluation Summary:

   This paper will be of broad interest to researchers interested in learning, memory, and/or the hippocampus. It offers a neuroanatomically inspired model of the hippocampus that reconciles its well-known role in episodic memory with its more recently appreciated role in category learning and generalization. The computational simulations are well conducted and support the key conclusions regarding complementary roles of distinct hippocampal pathways for different forms of learning. There are concerns with differentiating the current work from prior reports and the apparent discrepancy between the proposed model and well-established findings of place and concept cell recordings in hippocampus, but thought that these issues could be potentially resolved with additional clarification.

   (This preprint has been reviewed by …

   Evaluation Summary:

   This paper will be of broad interest to researchers interested in learning, memory, and/or the hippocampus. It offers a neuroanatomically inspired model of the hippocampus that reconciles its well-known role in episodic memory with its more recently appreciated role in category learning and generalization. The computational simulations are well conducted and support the key conclusions regarding complementary roles of distinct hippocampal pathways for different forms of learning. There are concerns with differentiating the current work from prior reports and the apparent discrepancy between the proposed model and well-established findings of place and concept cell recordings in hippocampus, but thought that these issues could be potentially resolved with additional clarification.

   (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. The reviewers remained anonymous to the authors.)

   Read the original source
5. 

   eLife

   Apr 26, 2022

   Reviewer #1 (Public Review):

   Sučević and Schapiro investigated a neurobiologically inspired model of human hippocampal structure and computation in category learning. In three separate simulations, the model (C-HORSE) is presented with learning tasks defined by various category structures from prior work and evaluated for its ability to learn the category structure, generalize categorization to novel stimuli, and accurately recognize previously encountered stimuli. Although originally conceived of as a computational model of associative memory, C-HORSE is demonstrated to quite naturally account for human-like learning of the three category tasks. Notably, the authors characterize the mechanisms underlying the model's learning by way of additional simulations in which "lesions" to the model's monosynaptic pathway (MSP; direct connections …

   Reviewer #1 (Public Review):

   Sučević and Schapiro investigated a neurobiologically inspired model of human hippocampal structure and computation in category learning. In three separate simulations, the model (C-HORSE) is presented with learning tasks defined by various category structures from prior work and evaluated for its ability to learn the category structure, generalize categorization to novel stimuli, and accurately recognize previously encountered stimuli. Although originally conceived of as a computational model of associative memory, C-HORSE is demonstrated to quite naturally account for human-like learning of the three category tasks. Notably, the authors characterize the mechanisms underlying the model's learning by way of additional simulations in which "lesions" to the model's monosynaptic pathway (MSP; direct connections between ERC and CA1) are contrasted with lesions to its trisynaptic pathway (TSP; pathway connecting ERC-DG-CA3-CA1). These in silico lesions offer key insight into the computational principles underlying theorized hippocampal functions in category learning: whereas MSP provides incremental learning of shared features diagnostic to category membership that are important for category generalization, TSP learns item-specific information that drives recognition behaviour. The authors propose that C-HORSE's successful account of a broad set of category learning datasets provides clear support for the role of complementary hippocampal functions mediated by MSP and TSP in category learning. This work adds compelling computational evidence to a growing literature linking hippocampus to a broader role in cognition that extends beyond declarative memory.

   The model simulations are clear and properly conducted. The three datasets examined offer a relatively broad set of findings from the category learning literature; that the models provide reasonable accounts of human performance in all three speaks to the model's generalizability. Overall, I find this work exciting and an important step in linking longstanding well-established formal learning theories of psychology with neurobiological mechanism. Several weaknesses dampen this excitement, each of which are detailed below:

   1. C-HORSE is presented as a new entry into a rich field of formal computational models of category learning. As noted above, the datasets examined span a broad range of learning contexts and structures and the model's ability to account for learning behaviour is compelling. However, no other models are leveraged to perform a direct evaluation. In other words, C-HORSE's predictions are compelling, but is it better than other competing models in the literature? To be clear, C-HORSE offers a novel alternative with its fundamental mechanisms originating from anatomical structure and connectivity. As such, a proof-of-concept showing that such a neurobiologically inspired framework can account for category learning behaviour is a worthwhile contribution in its own right and a clear strength of this paper. However, how to consider this model relative to existing theoretical frameworks is not well described in the manuscript.

   2. Relatedly, C-HORSE is evaluated in terms of qualitative fit to behaviour measures from prior studies and in all three simulations restricted to measure of end of learning performance. Again, an appeal to the proof-of-concept nature of the current work may provide an appropriate context for this paper. But, a hallmark of well-established category learning models (e.g., SUSTAIN, DIVA, EBRW, SEA, etc.) is their ability to account for both end of learning generalization (and in some cases, recognition) and behaviour throughout the learning process. C-HORSE does provide predictions of how learning unfolds over time, but how well this compares to human measures is not considered in the current manuscript. Such comparisons would strengthen the support for C-HORSE as a viable model of category learning and help position it in the busy field of related formal models.

   3. A consistent finding across all three simulations is that the TSP provides item-specific encoding. Evidence for this can be inferred by contrasting categorization and recognition performance across the TSP- and MSP-only model variants. In the discussion, the authors draw a parallel between exemplar theories of category learning and the TSP, which is a compelling theoretical position. However, as noted by the authors, unlike exemplar theories, the TSP-only model was notably impaired at categorization. The author's suggestions for extensions to C-HORSE that would enable better TSP-based categorization are interesting. But, I think it would be helpful to understand something about the nature of the representations being formed in the TSP-only model. For example, are they truly item-specific, are the shared category features simply lost to heightened encoding of item-unique features, are category members organized similarly to the intact model just with more variability, and so on. Characterizing the nature of these representations to understand the limitations of the TSP-only model seems important to understanding the representational dynamics of C-HORSE, but are not included in the current manuscript.

   4. In general, a detailed description that links model mechanisms and analyses to the learning constructs of interest for the different simulations is lacking. For example, RSA results for simulation 1 are contrasted for initial and settled representations, but what is meaningful about these two timepoints is not directly stated (moreover, what initial and settled response mean in terms of the current model is not explained). The authors do briefly suggest that differences between initial and settled representations may reflect encoding dynamics before and after big-loop recurrence, but this is not established as a key metric for evaluating the nature of the model representations. In general, more motivation is needed to understand what the chosen analyses reveal about the nature of the model's learning process and representations.

   5. I appreciate the comparison in the discussion to extant models of categorization. Certainly, the exemplar and prototype models are fixtures of the category learning literature and they somewhat align with the type of learning that TSP and MSP, respectively, provide. REMERGE and SUSTAIN are also briefly mentioned, but their discussion is limited which is unfortunate as they are actually more functionally equivalent to C-HORSE. I think, however, that the authors are missing an opportunity to discuss how C-HORSE offers a means for bridging levels of analysis to connect neurobiological mechanisms with these notably successful psychological models of category learning. Rather than framing C-HORSE as a competitor to existing models, it should be viewed as an account existing on a different level of analysis. In this sense, it complements existing approaches and potentially extends a theoretical olive branch between the psychology and neuroscience of category learning.

   6. The discussion takes a broad perspective on covering evidence concerning hippocampal contributions to category learning. Although comprehensive, some sections are not well connected back to the main thrust of the paper. For example, a section on neuropsychological accounts of the hippocampus and category learning summarizes central aspects of this literature but is never reflected on through the lens of the current findings. I do think this prior work is relevant, especially since it a central theme of the hippocampus not being necessary for category/concept learning, but its connection back to the current study is not well argued. Similarly, the section on consolidation and sleep is relevant, but in its current form does not seem to fit with the rest of the paper.

   Read the original source
6. 

   eLife

   Apr 26, 2022

   Reviewer #2 (Public Review):

   The authors present a model of the hippocampal region that incorporates both the (indirect) tri-synaptic and (direct) mono-synaptic pathways from entorhinal cortex (EC) to CA1 - the former incorporating projections from EC to dentate gyrus (DG), DG to CA3, and CA3 to CA1, and exhibiting a higher learning rate. They demonstrate that exposing this network to stimuli consistent with standard empirical tests of category learning (e.g. where within-category exemplars share a set of common features) allows the network to reliably assign both novel and previously encountered stimuli to the correct category (e.g. the network can learn to classify stimuli and generalise this knowledge to new examples). They show that the tri-synaptic pathway (TSP) preferentially supports the encoding of individual exemplars (e.g. …

   Reviewer #2 (Public Review):

   The authors present a model of the hippocampal region that incorporates both the (indirect) tri-synaptic and (direct) mono-synaptic pathways from entorhinal cortex (EC) to CA1 - the former incorporating projections from EC to dentate gyrus (DG), DG to CA3, and CA3 to CA1, and exhibiting a higher learning rate. They demonstrate that exposing this network to stimuli consistent with standard empirical tests of category learning (e.g. where within-category exemplars share a set of common features) allows the network to reliably assign both novel and previously encountered stimuli to the correct category (e.g. the network can learn to classify stimuli and generalise this knowledge to new examples). They show that the tri-synaptic pathway (TSP) preferentially supports the encoding of individual exemplars (e.g. analogous to episodic memory) while the mono-synaptic pathway (MSP) preferentially supports category learning.

   The manuscript is well written, the simulation details appear sound, and the results are clearly and accurately presented. This model builds on a long tradition of computational modelling of hippocampal contributions to human memory function, strongly grounded in anatomical and electrophysiology data from both rodents and humans, and is therefore able to link phenomena at the level of individual cells and circuits to emergent behaviour - a major strength of this, and similar, work. However, I have two major concerns relating to the relationship between these findings and previously published work by the same and other authors.

   First, it is not clear to me - from the manuscript - whether these results represent a significant novel advance on previous publications from the same senior author. Figures 1 and 3D are almost identical to figures published in Schapiro et al. (2017) Phil Trans B, and the take-home message (that the MSP might support statistical learning) is the same. In brief, it seems that the authors have subjected an identical network to some new (but related) tasks and reached the same set of conclusions. I see no distinction between learning to extract 'statistical regularities' (in previous work) and learning 'the structure of new categories' (described here). As an aside, demonstrating that an autoencoder network can learn stimulus categories and generalise to new exemplars is also well established.

   Second, I have some concerns with the relationship between the properties of this hippocampal network model and well described properties of single cells in the rodent and human hippocampus. In particular, the CA1 units in this model (and to some extent, also the CA3 units) come to respond strongly to all exemplars from within each category (e.g. as shown in Figure 3D, bottom right panel). This appears to be at odds with the known properties of place and concept cells from the rodent and human hippocampus, respectively, which show little generalisation across related concepts (i.e. the Jennifer Aniston neuron does not fire in response to other actors from Friends, for example). If the emergent properties of this model are not consistent with existing data, then it is not a valid model.

   More generally, the authors are clear that this model is "a microcosm of [the] hippocampus-neocortex relationship" and that the properties of the MSP "mirror those of neocortex". Why not assume that category learning is supported by an interaction between hippocampus and neocortex, then, as in the complementary learning systems (CLS) model? Aside from some correlational fMRI data and partial deficits in hippocampal amnesics - either of which could have a myriad of different explanations - what empirical data is better accounted for by this model than CLS? Put differently, what grounds are there for rejecting the CLS model? To some extent, this model appears to account for less empirical data than CLS, with the exception of a few recent neuroimaging studies (which are hard to interpret at the level of single cells)

   Read the original source
7. 

   eLife

   Apr 26, 2022

   Reviewer #3 (Public Review):

   The current work aimed to determine how the hippocampus may be able to detect regularities across experiences and how such a mechanism may serve to support category learning and generalization. Rapid learning in the hippocampus is critical for episodic memory and encoding of individual episodes. However, the rapid binding of arbitrary associations and one-shot learning was long thought suboptimal for finding regularities across experiences to support generalization, which were instead ascribed to other, slower-learning memory systems. More recent work has started to highlight hippocampal role in generalization, renewing the question of how generalization can be accomplished alongside memory for episodic details within a single memory structure. The current paper offers a reconciliation, presenting a …

   Reviewer #3 (Public Review):

   The current work aimed to determine how the hippocampus may be able to detect regularities across experiences and how such a mechanism may serve to support category learning and generalization. Rapid learning in the hippocampus is critical for episodic memory and encoding of individual episodes. However, the rapid binding of arbitrary associations and one-shot learning was long thought suboptimal for finding regularities across experiences to support generalization, which were instead ascribed to other, slower-learning memory systems. More recent work has started to highlight hippocampal role in generalization, renewing the question of how generalization can be accomplished alongside memory for episodic details within a single memory structure. The current paper offers a reconciliation, presenting a biologically-inspired model of the hippocampus that is able to learn categories alongside stimulus-specific information comparably to human performance. The results convincingly demonstrate how distinct pathways within the hippocampus may differentially serve these complementary memory functions, enabling the single structure to support both episodic memory and categorization.

   Major strengths and contributions

   The paper includes simulation of three distinct categorization tasks, with a clear explanation of the unique aspects of each task. The key results are consistent across tasks, lending further support to the main conclusions of the role of distinct hippocampal pathways in learning specific details vs. regularities. Together with prior work on how the same architecture can support statistical learning in other types of tasks, this work provides important evidence of the broad role of the hippocampus in rapid integration of related information to serve many forms of cognition.

   Throughout the paper, the authors nicely explain in conceptual terms how the same underlying computations may serve all three categorization tasks as well as statistical learning and episodic inference tasks. Thus, the paper will be of broad interest, beyond researchers focused on modeling and/or categorization.

   On a conceptual level, this work provides a fruitful framework for understanding hippocampal functions, representations and computations. It provides a highly plausible mechanistic explanation of how category learning and generalization can be accomplished in the hippocampus and how distinct types of representations may emerge in distinct hippocampal subfields. The framework can be used to derive new testable predictions, some of which the authors themselves introduced. It also provides new insights into how the outputs of different pathways influence each other, providing a more nuanced view of the division of labor and interactions between hippocampal subfields. For example, the big loop recurrence would eventually lead to category influences even on the initially sparse, pattern separated representations in the CA3, which is an idea consistent with empirical observations.

   The presented computational model of the hippocampus is currently the most detailed and biologically plausible hippocampal model easily applicable in the area of cognitive neuroscience and behavioral simulations. The commonalities and differences with other related models (conceptual and computational) are well explained. Both the conceptual and technical descriptions of the model are exceptionally clear and detailed. The model is also publicly available for download for any researcher to use with their own task and data. All these aspects make it likely that other researchers may adopt the model in a wider range of tasks, stimulating new discoveries.

   The autoencoder nature of the model and the use of categorization tasks meant that some measures of interest, like recognition of exemplar-specific information, could not be evaluated by direct reading of the output layer to compare with some label (like old/new). The authors however came up with clever ways how to evaluate recognition performance in each task that was sensible and highlighted the multiple ways how one may think about information contained in neural representations in each layer. This approach can also be utilized by others for evaluating item-specific and category information in activation patterns, for example in analyses of fMRI.

   Finally, I thought the current paper and provided model may also serve as an excellent introduction to computational modeling for those new to this approach. The exceptional clarity of the conceptual and technical description of this model and the clear logic of how one may model a cognitive task and interpret results made this paper fairly accessible. Furthermore, the paper offered new insights and predictions based on analyzing the model's hidden layers, lesion performance, and/or noting some patterns of behavior unique to specific tasks. This was also instructive for highlighting the distinctive contributions that the computational modeling approach can have for furthering our understanding of cognition and the brain.

   Weaknesses

   The paper's strengths far outnumbered the weaknesses, that are minor. For one, the selected categorization tasks nicely complemented each other, but only covered stimuli with discrete-value dimensions (features like color, shape, symbol, etc). The degree to which the results generalize (or not) to continuous-value stimuli and different category structures (for instance information-integration or rule-based in COVIS framework) is not clear. How the model could be adjusted for continuous-value stimuli was not specified.

   There is compelling evidence for the dissociation between different hippocampal pathways and subfields (CA1 vs. CA3) that the model is based on. As the authors noted, there is also compelling evidence for functional dissociations along the long hippocampal axis, with anterior portions more geared towards coarse, generalized representations while posterior towards more detailed, specific representations. The authors nicely pointed out that these proposals of within-hippocampus division of labor are less orthogonal than they may first appear, as there is greater proportion of CA1 in the anterior hippocampus. However, it is premature to imply that this resolves the CA1/CA3 vs. anterior/posterior question; the idea that existing anterior findings may be simply CA1 findings is currently only speculation. Furthermore, first studies indicating that anterior/posterior representational gradients may exist within each subfield are beginning to emerge.

   Read the original source
8. 

   Version published to 10.1101/2022.01.12.476051v1 on bioRxiv

   Jan 13, 2022