acceptodds
Under review as a conference paper at ICLR 2027

Cognitive Map as Reference Path Measure: The Hippocampus as a Schr\"odinger Bridge Sampler

Abstract

A cognitive map enables animals to recall past journeys, plan future routes, and imagine unexperienced paths. How a single neural circuit generates such diverse sequential experiences from learned knowledge remains a fundamental open question. We propose that a cognitive map defines a reference path measure, a probability distribution over sequence spaces, and that the hippocampus constructs goal-directed episodes by solving a Schr\"odinger bridge problem under endpoint constraints. In this framework, present cues and terminal evidence modulate sequence generation through a continuation field that evaluates path reachability to target outcomes. We demonstrate that this computation is naturally realized by hippocampal circuit dynamics: recurrent relaxation computes the continuation field, dendritic gain modulates assembly transitions, and inhibitory competition governs event completion. Furthermore, offline replay supplies internal endpoint expectations, enabling contrastive calibration of boundary representations against waking experience. Across network simulations and large-scale hippocampal electrophysiological recordings (HC-3, HC-11, and DANDI 000978), our model recovers exact path laws, reveals reusable temporal structure, accurately predicts complete-event trajectories and durations under boundary conditioning, and accounts for experience-dependent shifts in sleep replay. This work unifies attractor dynamics with conditional generative modeling, providing a shared computational foundation for memory retrieval, prospective planning, and offline consolidation.

Primary area:
applications to neuroscience & cognitive science
Keywords:
cognitive map, Schr\"odinger Bridge,Hippocampus,Attractor Dynamics,Generative Modeling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_cJy2OpenReview ↗
Rating
1
Confidence
2(+1)
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes that a cognitive map defines a probability distribution over sequence spaces, and that the hippocampus function can be understood as solving a Schrödinger bridge problem. The model is evaluated on publicly available hippocampal electrophysiological recordings.

Critical strengths

Connecting the hippocampus to Schrödinger bridge may be an interesting idea. Unfortunately, in its current form, the paper does not allow for a clear assessment of this contribution.

Critical weaknesses

The current presentation of the paper prevents an accurate assessment of the paper’s claim. Many paragraphs present several claims with little explanation or motivation. Key details are scattered across the 121 pages, making it difficult to follow the core arguments. The novelty is unclear. For example, Piray & Daw (2021) proposed a similar model also aiming to better understand the reusable cognitive map of the hippocampal-entorhinal system. However, this work is only cited once in Appendix B.6 at page 46 as “Related positive transforms support linearly solvable control (Todorov, 2009) and reusable planning in cognitive maps (Piray & Daw, 2021; 2025).”.

Some of the paper's ideas may be interesting, but it would require a substantial revision to determine it.

Questions for authors

None

Reviewer_qEzbOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Summary of contributions:

  • The authors propose a computational framework for understanding how an internal hippocampal map can support memory retrieval, planning, and offline replay.
  • Starting from an available map of transitions between represented situations, the authors model how information about an episode’s beginning and ending can guide the generation of a complete sequence.
  • The authors support the framework through network simulations, and comparisons with existing hippocampal recordings. They also propose a learning rule that adjusts which beginnings and endings the model favors after experience.
Critical strengths

Strengths:

  • Connecting memory and computation: I find the attempt to connect retrieval, planning, and replay within a shared framework interesting. The proposal offers a way to understand how the same internal map could support different uses of previously acquired knowledge.
  • Connecting theory with neural observations: Combining theoretical analysis, simulations, and electrophysiological recordings is a strength. The simulations demonstrate the proposed computation, while the recording analyses examine whether it can account for observed features of hippocampal sequences and replay.
Critical weaknesses

Weaknesses:

  • Clarity: I do not work directly in this field, and I found the paper difficult to follow despite careful reading. I appreciate the authors’ efforts to explain the ideas to a broader audience, but a simpler overview would help clarify what the model assumes, what it learns, and what it predicts. It would also be useful to distinguish more clearly which aspects of the broad account are directly supported by the experiments and which remain proposed interpretations.
  • Relationship to prior work: I found it difficult to identify the precise advance over existing models. Much of the relevant discussion appears in the appendix, while the main text emphasizes the broad framework. A concise main-text comparison explaining what earlier approaches already account for, what this proposal adds, and which results demonstrate that added value would help readers outside the field assess the contribution.
Questions for authors

NA.

Reviewer_Nu2SOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposed Schroedinger bridge as a hippocampus sampler, particularly for planning and navigation in cognitive maps. The authors propose a recurrent circuit that implements the identified rules, and report experimental results from this model in 4 different experiments.

Critical strengths
  • Clear formalism
  • Clear and falsifiable predictions
Critical weaknesses
  • Only weak experimental evidence in support of the formalism
Questions for authors
  • Are there applications of the identified principles in ML/DL?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.