Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 38 of the 38 references it cites were found in a registry or an index; 3 of the 8 sources its claims cite were checked for support.

Unfolding the Black Box of Recurrent Neural Networks for Path Integration

Open access
Chu, Tianhao, Wu, Yuling, Burgess, Neil, Ji, Zilong, Wu, Si · 2025 · Advances in Neural Information Processing Systems 38

Abstract

Abstract Path integration is essential for spatial navigation. Experimental studies have identified neural correlates for path integration, but exactly how the neural system accomplishes this computation remains unresolved. Here, we adopt recurrent neural networks (RNNs) trained to perform a path integration task to explore this issue. After training, we borrow neuroscience prior knowledge and methods to unfold the black box of the trained model, including: clarifying neuron types based on their receptive fields, dissecting information flows between neuron groups by pruning their connections, and analyzing internal dynamics of neuron groups using the attractor framework. Intriguingly, we uncover a hierarchical information processing pathway embedded in the RNN model, along which velocity information of an agent is first forwarded to band cells, band and grid cells then coordinate to carry out path integration, and finally grid cells output the agent location. Inspired by the RNN-based study, we construct a neural circuit model, in which band cells form one-dimensional (1D) continuous attractor neural networks (CANNs) and serve as upstream neurons to support downstream grid cells to carry out path integration in the 2D space. Our study challenges the conventional view of considering grid cells as the principal velocity integrator, and supports a neural circuit model with the hierarchy of band and grid cells.

Overview

CoChat Summary

TL;DR

The authors train recurrent neural networks on a path integration task and then apply neuroscience-style analyses (receptive-field classification, connection pruning, attractor dynamics) to reveal a hierarchical pathway in which velocity input goes to band cells, band and grid cells interact reciprocally, and grid cells output position. Inspired by this, they build a hand-designed circuit model of 1D band-cell CANNs feeding a 2D grid-cell CANN, arguing against the conventional view that grid cells alone are the principal velocity integrator.

Key findings

  • In one trained RNN with N_R = 4096 units, the authors identify 764 grid cells, 764 band cells, and 2568 undefined cells; all three types emerge robustly across seeds.
  • Pruning velocity inputs to band cells produces the largest error increase (e.g., 2.32 ± 0.43% vs 0.11 ± 0.07% for grid cells and 0.14 ± 0.09% for undefined cells when 750 connections are pruned), while pruning read-out connections from grid cells has the largest effect (9.15 ± 0.71% vs 2.67 ± 0.50% for band cells).
  • Pruning either band→grid or grid→band connections substantially degrades performance, indicating bidirectional cooperation is critical; a new RNN built with the hierarchical pathway as a prior matches the original's performance and develops band and grid cells in the first and second modules respectively, while a location-reconstruction task yields grid cells but no band cells.
  • Band cells cluster into four groups by spacing and orientation, show direction tuning with preferred directions at θ or θ + π, form ring-like Isomap embeddings consistent with 1D CANNs, and have center-surround-like connectivity; the proposed circuit requires 2N pure band + 4N conjunctive band + N² grid cells versus 5N² for grid-only models.

Methods

  • Vanilla RNN (N_R = 4096, ReLU, r_{t+1} = σ(Jr_t + Mv_t), read-out p_t = Wr_t) trained by BPTT with cross-entropy loss on a path integration task in a 2.2 m × 2.2 m arena, initialized from 512 place cells (difference-of-softmax tuning), 100 epochs, RMSProp, learning rate 10⁻⁴.
  • Neuron types classified from spatial rate maps using spectral band score (>5.8) and grid score (>0.88) criteria; four step-by-step pruning experiments (velocity inputs, read-out connections, cell types, band–grid connections) with error bars over n = 30 repetitions.
  • Population dynamics analyzed via PCA (15 components) followed by Isomap (3D, k = 8); a hand-designed neural circuit model with 1D band-cell CANNs and a 2D grid-cell CANN was simulated over 100,000 steps (Δt = 0.05) for activity mapping and 1,000 steps for trajectory decoding.

Limitations

  • Stated by the authors: it is unclear whether undefined cells have spatial tuning (e.g., head direction or boundary vector properties) or contribute to path integration; the nature of band–grid interactions in the trained RNN is not fully solved and may not map directly onto the proposed circuit's connectivity; and the training paradigm is restricted to physical space, leaving generalization to abstract spaces untested.
  • Evident in the text: the main circuit-model simulation is idealized with no noise in inputs or model (noise-induced drift is deferred to SI), and the authors report no compute resources, stating such specifications were deemed nonessential.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.52202/085713-3412

Cites

38 of 38 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

3 Valid · 0 Not valid · 4 Not checked · 59 Own results and interpretations

36 more claims

Create a free account to see all 66 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 15
  • Probably not a substantial claim. 12
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Show all 29 references

Research works printed but not linked: 1

  1. [29]29. Zilong Ji, Tianhao Chu, Xingsi Dong, Changmin Yu, Daniel Bush, Neil Burgess, and Si Wu. Dynamical modulation of hippocampal replay sequences through firing rate adaptation. BioRxiv, pages 2024-09, 2024.

Cited by

See all 3

Metadata

Authors
Chu, T., Wu, Y., Burgess, N., Ji, Z., & Wu, S.
Affiliations
  • Peking University (CN)
  • Peking University
  • University College London (GB)
Corresponding author
Wu, Si
Venue
Advances in Neural Information Processing Systems 38
Published
2025
Pages
113045-113078
ISSN
1049-5258 (ISSN-L)
OpenAlex
W4415559134
Cited by
3
References
29
Citations by year
2026: 3
FWCI
6.72
Type
conference paper
Language
en
Open access
Licence
cc-by
Locations
  • neural information processing systems (publishedVersion)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by)
Funders
  • National Natural Science Foundation of China (T2421004, 2021ZD0200204)
  • National Key Research and Development Program of China (2021ZD0200204, 2024YFF1206500)
Topics
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Memory and Neural Mechanisms
  • Social Sciences › Psychology › Experimental and Cognitive Psychology › Spatial Cognition and Navigation
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Neural dynamics and brain function
Keywords
path integration, recurrent neural networks, spatial navigation, grid cells, continuous attractor networks, neural circuit model, hierarchical information processing, velocity integration, attractor dynamics, receptive fields, connection pruning

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-01, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-01 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Unfolding the Black Box of Recurrent Neural Networks for Path Integration. https://app.cochat.ai/papers/10.52202/085713-3412
The paper
Chu, T., Wu, Y., Burgess, N., Ji, Z., & Wu, S. (2025). Unfolding the Black Box of Recurrent Neural Networks for Path Integration. Advances in Neural Information Processing Systems 38. https://doi.org/10.52202/085713-3412

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.