The Tolman-Eichenbaum Machine: Unifying Space and Relational Memory through Generalization in the Hippocampal Formation

Abstract

Summary The hippocampal-entorhinal system is important for spatial and relational memory tasks. We formally link these domains, provide a mechanistic understanding of the hippocampal role in generalization, and offer unifying principles underlying many entorhinal and hippocampal cell types. We propose medial entorhinal cells form a basis describing structural knowledge, and hippocampal cells link this basis with sensory representations. Adopting these principles, we introduce the Tolman-Eichenbaum machine (TEM). After learning, TEM entorhinal cells display diverse properties resembling apparently bespoke spatial responses, such as grid, band, border, and object-vector cells. TEM hippocampal cells include place and landmark cells that remap between environments. Crucially, TEM also aligns with empirically recorded representations in complex non-spatial tasks. TEM also generates predictions that hippocampal remapping is not random as previously believed; rather, structural knowledge is preserved across environments. We confirm this structural transfer over remapping in simultaneously recorded place and grid cells.

Topographic connectivity reveals task-dependent retinotopic processing throughout the human brain

“…The present finding, that the hippocampus shares a retinotopic mode of organization with much of the rest of the brain, is in line with canonical tracer-based network findings ( 18 ). As hippocampus also entertains world-centric coding of space ( 34 ), this solidifies the notion ( 19 ) that hippocampus is a nexus for the conjunctive coding of both world-centric and sensory reference frames ( 20 , 35 ).…”

“…Although it is not generally implicated in traditional vision science experiments, tracer-based connectivity studies place the hippocampal formation at the top of the visual-processing hierarchy ( 18 ). Hippocampus is thought to implement the interaction between memory-related and sensory processing or imagery ( 19 – 22 ), leading me to reason that both narrative understanding of naturalistic inputs and internally generated thought should evoke strong RC in hippocampus, over and above its recently discovered contralateral visual field preference during retinotopic mapping ( 23 ) ( SI Appendix , Fig. S4 ).…”

When to retrieve and encode episodic memories: a neural network model of hippocampal-cortical interaction

“…Hulbert and Norman, 2015; Kim et al, 2017; Ritvo et al, 2019; Schapiro et al, 2012; Schapiro et al, 2016; M. L. Schlichting et al, 2015; Stachenfeld et al, 2017; Whittington et al, 2020). Similarly, using a more biologically detailed cortical model (separated into distinct cortical sub-regions) could help us to connect to data on how different cortical regions interact with hippocampus during event processing (e.g., Barnett et al, 2020; Cooper et al, 2020; Gilboa and Marlatte, 2017; Preston and Eichenbaum, 2013; Ranganath and Ritchey, 2012; Ritchey and Cooper, 2020; van Kesteren et al, 2012).…”

Inferences on a Multidimensional Social Hierarchy Use a Grid-like Code

“…), these cells may provide a neuronal basis for a cognitive map of space. Critically, recent biologically inspired computational models have further proposed how these grid codes in EC may reflect a basis for statistical transitions in 2-D topologies that can in principle enable animals and artificial agents to compose new routes and find shortcuts to reach goals in spatial environments (Bush et al, 2015;Banino et al, 2018;Behrens et al, 2018;Whittington et al, 2020). These findings suggest a novel framework for understanding how humans might effectively "shortcut" solutions in new situations.…”

“…Representing non-spatial relationships between entities or task states in the world as a cognitive map would likewise be incredibly powerful because it could theoretically be leveraged to make inferences from sparse experiences that can dramatically accelerate learning and even guide decisions never faced before (Kriete et al, 2013;Behrens et al, 2018;Wang et al, 2018). This is because, in theory, a cognitive map of abstract relationships, such as a conceptual space (Constantinescu, O'Reilly and Behrens, 2016), or task space (Schuck et al, 2016;Zhou et al, 2019), would allow for "novel routes" and "shortcuts" to be inferred, as in physical space (Behrens et al, 2018;Whittington et al, 2020) (Fig.1a). Notably, such direct inferences go beyond chaining together directly experienced associations (Fig.1a), and are seen as the defining characteristic of a genuine cognitive map of spatial environments (Tolman, 1948;O'Keefe and Nadel, 1978;Bennett, 1996).…”