ReGraph: A Computational Account of Emergent Generalization in the "what" and "where" Dual Visual Streams

Hyewon Kang; Jungmin Lee; Ilgyu Lee; Seok-Jun Hong

cs.NEOct 6, 2026

ReGraph: A Computational Account of Emergent Generalization in the "what" and "where" Dual Visual Streams

Authors: Hyewon Kang, Jungmin Lee, Ilgyu Lee, Seok-Jun Hong

Organizations: Department of Intelligent Precision Healthcare Convergence, Sungkyunkwan University, Suwon, Republic of Korea · Center for Neuroscience Imaging Research, Institute for Basic Science (IBS), Suwon, Republic of Korea · Department of Biomedical Engineering, Sungkyunkwan University, Suwon, Republic of Korea

Abstract

Where generalization capacity--the ability to extract context-invariant relational structures--first emerges remains a central question in AI and neuroscience. The foundation for this capacity lies upstream of the hippocampus, within the entorhinal cortex, where parallel pathways dissociate relational structure in the medial entorhinal cortex (MEC) from sensory content in the lateral entorhinal cortex. However, as Eichenbaum argued, such factorization likely originates earlier, driven by the segregation of the dorsal ('where') and ventral ('what') visual streams. Supporting this, grid-like firing patterns--a signature of MEC (context-invariant codes)--also appear in preceding neocortical regions along the dorsal pathway. Yet, how such representations are computationally formed along upstream pathways remains unknown. To investigate this in silico, we developed ReGraph, a recurrent dual-stream graph model with biological inductive biases, including retina-driven stream-specialized encoding, dorsal-to-ventral modulation, and dynamic lateral connectivity. Trained on the action benchmark Something-Something V2, ReGraph revealed a pathway-specific emergence of relational mapping: context-invariant codes and grid-like spatial bases uniquely co-emerged along the extended dorsal stream. In contrast, their absence in single-stream, unmodulated variants, and standard baselines implies that these inductive biases are prerequisites for relational structures. Crucially, our post-hoc analyses demonstrated that these grid-like bases serve as reusable routing templates for information processing via lateral connectivity. Together, our findings provide a computational account that generalization may not be a faculty that emerges abruptly within a dedicated region, but a property that already takes shape as sensory information is parsed into factorized streams of hierarchical visual processing.

Figures & tables

Figure 1: From retina to dual visual streams. Left : The retinal M/P dichotomy drives the dorsal and ventral pathways, each comprising primary and extended stages. Grid-like patterns in the extended dorsal pathway motivate a hypothesis that the precursor of relational structures may emerge along this pathway. The primary dorsal/ventral pathways preferentially encode spatiotemporal dynamics and object details, respectively. Right: Schematics of three connectivity principles defined in § 2.2 .

Figure 2: The ReGraph architecture. A. Input is tokenized into M/P-driven asymmetric views (Dorsal: high temporal/low spatial; Ventral: low temporal/high spatial) and processed via specialized pretrained networks for spatiotemporal process (depth, optical flow, saliency) and hierarchical object-oriented processing. B. Stacked layers model cortical hierarchies, with top-down modulation routing the dorsal signals to the ventral layers. C. Self-attention acts as a dynamic graph operator, followed by a GRU-based recurrent update to regulate temporal flow.

Ablation Type	Variant	$\mathbf{T_{D}/T_{V}}$	$\mathbf{N_{D}/N_{V}}$	V-RGB	$\mathbf{D{\to}V}$	Tokens	$\begin{subarray}{c}\textbf{Top-1}\\ \textbf{(\%; ID)}\end{subarray}$
Single-Stream	Dorsal-Only	12 / –	196 / –	–	–	2,352	71.69
References	Ventral-Only	– / 3	– / 784	Yes	–	2,352	60.38
Dorsal–Ventral	(a) Trait-Symmetric	5 / 5	484 / 484	No	$\times$	4,840	69.78
Biological Trait	(b) Dorsal-trait Only	8 / 2	484 / 484	No	✓	4,840	72.42
Asymmetry	(c) Ventral-trait Only	5 / 5	196 / 784	Yes	$\times$	4,900	70.43
	(d) Fully Asymmetric (ReGraph)	12 / 3	196 / 784	Yes	✓	4,704	74.57

Table 1: Ablation over design choices. All conditions share a fixed token budget (

T_{D}N_{D}+T_{V}N_{V}\approx 4{,}800

). Temporal (

T_{D}\!>\!T_{V}

) and spatial (

N_{V}\!>\!N_{D}

) imbalances realize the proposed dorsal and ventral traits, respectively. ID: In-distribution evaluation.

Model	Stream	L1 (OOD [Drop])	L2 (OOD [Drop])	L3 (OOD [Drop])	L4 (OOD [Drop])	Avg (OOD [Drop])
Dorsal-Only	Dorsal	72.8 $\pm$ 3.6 [ $-$ 25.6]	69.0 $\pm$ 3.7 [ $-$ 30.7]	76.7 $\pm$ 1.9 [ $-$ 23.2]	93.6 $\pm$ 1.6 [ $-$ 6.4]	78.0 [ $-$ 21.5]
Unmodulated	Dorsal	86.8 $\pm$ 2.7 [ $-$ 12.7]	82.9 $\pm$ 2.1 [ $-$ 17.1]	81.5 $\pm$ 2.1 [ $-$ 18.5]	96.9 $\pm$ 0.4 [ $-$ 3.1]	87.0 [ $-$ 12.9]
Unmodulated	Ventral	52.0 $\pm$ 3.8 [ $-$ 44.2]	56.9 $\pm$ 2.7 [ $-$ 42.1]	59.2 $\pm$ 4.5 [ $-$ 40.2]	65.4 $\pm$ 1.7 [ $-$ 33.4]	58.4 [ $-$ 40.0]
ReGraph	Dorsal	82.0 $\pm$ 2.7 [ $-$ 16.9]	81.5 $\pm$ 3.2 [ $-$ 18.5]	78.5 $\pm$ 3.4 [ $-$ 21.3]	97.4 $\pm$ 0.6 [ $-$ 2.6]	84.9 [ $-$ 14.8]
ReGraph	Ventral	55.0 $\pm$ 5.7 [ $-$ 43.6]	59.7 $\pm$ 1.7 [ $-$ 40.0]	95.9 $\pm$ 1.1 [ $-$ 4.0]	95.0 $\pm$ 2.5 [ $-$ 4.9]	76.4 [ $-$ 23.1]

Table 2: OOD linear probing for context-invariance. OOD [Drop] : OOD accuracy (%) and its train–test gap (pp), averaged over 8 heads (30 samples/class). Bold: superior OOD accuracy in the dorsal streams (avg.

>

80%) and the L3–L4 surge in the ventral stream of ReGraph (

D{\to}V

Model	L1	L2	L3	L4	Avg
Dorsal-Only	0.0 / –	5.6 / –	4.0 / –	2.6 / –	3.0 / –
Unmodulated	0.1 / 0.1	8.6 / 6.8	4.0 / 5.9	6.3 / 5.9	4.7 / 4.7
ReGraph	0.0 / 0.1	5.1 / 6.5	8.8 / 6.0	15.3 / 7.2	7.3 / 4.9

Table 3: Ratio of grid-like bases across layers. We report the fraction of top 5% singular-value bases exceeding a shuffled null threshold (%), formatted as Dorsal / Ventral (App. J ). Bold indicates a strictly increasing trend unique to ReGraph ’s extended dorsal stream.

Figure 3: Spatial weighting maps in ReGraph . 2D autocorrelograms of pre-softmax attention bases; GS: gridness score. Dorsal (left three) : hexagonal patterns with varied orientation and spacing. Ventral (right three) : irregular or cross-like patterns lacking hexagonal structure, with near-zero GS.

Task		L1 ( $G_{\text{high}}/G_{\text{low}}[\Delta\uparrow]$ )	L2 ( $G_{\text{high}}/G_{\text{low}}[\Delta\uparrow]$ )	L3 ( $G_{\text{high}}/G_{\text{low}}[\Delta\uparrow]$ )	L4 ( $G_{\text{high}}/G_{\text{low}}[\Delta\uparrow]$ )
Bases Ablation	$\sigma=0$	82.0 / 81.7 [ $-0.3$ ]	82.6 / 82.9 [ $+0.3$ ]	70.3 / 77.3 [ +7.0 ]	89.1 / 95.1 [ +6.0 ]
	$\sigma=0.5$	70.3 / 70.1 [ $-0.2$ ]	75.3 / 75.4 [ $+0.2$ ]	61.7 / 70.2 [ +8.5 ]	85.8 / 93.6 [ +7.7 ]
	$\sigma=1.0$	54.3 / 54.4 [ $+0.1$ ]	60.9 / 60.2 [ $-0.7$ ]	47.1 / 55.8 [ +8.7 ]	75.4 / 86.9 [ +11.5 ]
	$\sigma=1.5$	42.0 / 42.0 [ $0.0$ ]	47.5 / 46.1 [ $-1.4$ ]	35.9 / 43.4 [ +7.5 ]	61.2 / 75.2 [ +14.0 ]
Reconstruction ( $R^{2}$ )		.000 / .014 [ $-.014$ ]	.008 / .001 [ $+.007$ ]	.082 / .000 [ +.082 ]	.131 / .037 [ +.094 ]

Table 4: Functional relevance of grid-like bases. Each cell reports results obtained under

G_{\text{high}}

G_{\text{low}}

manipulations [gap

\Delta_{Acc./R^{2}}\uparrow

]. A larger

\Delta

indicates greater reliance on

G_{\text{high}}

for both tasks (OOD Acc. under ablation; reconstruction score). Bold (

\Delta>\text{row mean}

) highlights the L3-4 surge.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	GNN role	Biological analog
$\psi(\cdot)$	Message function	Axonal/synaptic transmission gated by $\mathbf{e}_{ji}$
$\bigoplus$	Permutation-invariant aggregation	Dendritic integration of postsynaptic potentials
$\phi(\cdot)$	Update function	Non-linear somatic activation producing the next firing state

Appendix

Table 5: Correspondence between message-passing operations and cortical lateral processing.

General MP component	General form	GCN realization
Message $\psi$	$\psi(\mathbf{h}_{i}^{\ell},\mathbf{h}_{j}^{\ell},\mathbf{e}_{ji})$	$\dfrac{1}{\sqrt{\tilde{d}_{i}\tilde{d}_{j}}}\,\mathbf{W}\mathbf{h}_{j}^{\ell}$
Aggregation $\bigoplus$	permutation-invariant op	$\displaystyle\sum_{j\in\mathcal{N}_{i}\cup\{i\}}$
Update $\phi$	$\phi(\mathbf{h}_{i}^{\ell},\text{agg})$	$\sigma(\text{agg})$

Appendix

Table 6: GCN viewed as a specific instance of the general message-passing framework in Eq. equation 4 .

Functional role	GCN	Self-Attention
Adjacency (who connects to whom)	$\hat{\mathbf{A}}_{ij}$ (scalar, fixed)	$\dfrac{\mathbf{q}_{i}^{\top}\mathbf{k}_{j}}{\sqrt{d_{k}}}$ (input-dependent)
Edge-weight normalization	$1/\sqrt{\tilde{d}_{i}\tilde{d}_{j}}$ (degree-based)	softmax over $j$ (similarity-based)
Connectivity scope	Local: $j\in\mathcal{N}_{i}\cup\{i\}$	Global: every pair $(i,j)$ , $1\leq i,j\leq N$
Source of connectivity	Predefined graph topology	Learned via $\mathbf{W}_{Q},\mathbf{W}_{K}$
Temporal behavior	Fixed across all inputs	Recomputed for every $\mathbf{H}^{\ell}$
Transmitted signal	$\mathbf{W}\mathbf{h}_{j}^{\ell}$	$\mathbf{V}_{j}=\mathbf{W}_{V}\mathbf{h}_{j}^{\ell}$

Appendix

Table 7: Component-level equivalence between GCN and self-attention. Self-attention preserves the graph-operator structure of GCN but lifts the adjacency from a static, sparse object to a dynamic, dense one.

Stream	Cue	Model	Extraction Block	Output Dim.	Post-processing
Dorsal	Depth	Depth Anything V2	refinenet2	$64\times 16\times 16$	Flatten to $N_{D}{=}196$ , project to $d{=}512$ , fuse via cross-attention $\to\mathbf{E}_{D}$
	Flow	GMFlow	Conv1 (upsampler)	$256\times 28\times 28$
	Saliency	UniSal	upsample_2	$128\times 28\times 28$
Ventral	Object	ViT-S/8 (DINO)	Block 9	$384\times 28\times 28$	Discard [CLS] , project to $d{=}512\to\mathbf{E}_{V}$ ( $N_{V}{=}784$ )

Appendix

Table 8: Pretrained network specifications for the primary visual stage. “Extraction Block” denotes the specific layer from which representations are drawn. “Output Dim.” specifies the raw channel and spatial shape (

C^{\prime}\times H^{\prime}\times W^{\prime}

) per frame, prior to spatial resampling and linear projection that produce the tokenized form

\mathbb{R}^{N_{S}\times d}

ID	Action Class
6	Covering something with something
12	Dropping something onto something
15	Hitting something with something
16	Holding something
19	Holding something next to something
36	Moving something and something away from each other

Appendix

Table 9: The 33 action classes selected from SSV2, sorted by class ID. IDs correspond to the original 174-class indexing of the SSV2 dataset.

Figure 4: Example videos and corresponding templated descriptions ( Goyal et al., 2017 ) .

Hyperparameter	Single-Stream	Dual-Stream
Learning rate (ReGraph layers)	5e-4	3e-4
Learning rate (pretrained networks)	5e-5	3e-5
Weight decay	0.03	0.05
Batch size	5	5
Gradient accumulation	4	6
Effective batch size	20	30

Appendix

Table 10: Training hyperparameters. The ”Single-Stream” column applies to both single-stream references (Dorsal-Only and Ventral-Only), and the ”Dual-Stream” column applies to all dual-stream variants.

Layer	H1	H2	H3	H4	H5	H6	H7	H8	Mean ( $\pm$ std)
$V_{1}$	0.292	0.283	0.249	0.294	0.326	0.262	0.253	0.269	0.279 ( $\pm$ 0.024)
$V_{2}$	0.277	0.287	0.313	0.298	0.277	0.283	0.290	0.283	0.289 ( $\pm$ 0.011)
$V_{3}$	0.444	0.479	0.478	0.491	0.481	0.444	0.482	0.472	0.471 ( $\pm$ 0.017)

Appendix

Table 11: Learned

D{\to}V

gate values

\sigma(\gamma)

per head at each ventral layer of ReGraph .

V_{k}

(

k\in\{1,2,3\}

) denotes the

k

-th ventral layer;

V_{4}

is omitted as it receives no

D{\to}V

modulation. Bold highlights the sharp opening at

V_{3}

across all heads.

Model	Stream	L1	L2	L3	L4	Avg
Dorsal-Only	Dorsal	0 (0.0%)	105 (5.6%)	75 (4.0%)	48 (2.6%)	228 (3.0%)
Unmodulated	Dorsal	1 (0.1%)	161 (8.6%)	74 (4.0%)	118 (6.3%)	354 (4.7%)
Unmodulated	Ventral	9 (0.1%)	553 (6.8%)	479 (5.9%)	478 (5.9%)	1519 (4.7%)
ReGraph	Dorsal	0 (0.0%)	95 (5.1%)	164 (8.8%)	286 (15.3%)	545 (7.3%)
ReGraph	Ventral	9 (0.1%)	525 (6.5%)	483 (6.0%)	588 (7.2%)	1605 (4.9%)

Appendix

Table 12: Counts and ratios of grid-like spatial bases (Top-5% retention rate). Each cell shows

n_{\text{grid}}

( Ratio% ), aggregated across classes and heads. Bold highlights ReGraph’s dorsal stream, exhibiting monotonic emergence of grid-like bases.

Model	Stream	L1	L2	L3	L4	Avg
Dorsal-Only	Dorsal	.244 / –	.240 / .272	.244 / .324	.242 / .341	.243 / .312
Unmodulated	Dorsal	.242 / .269	.240 / .409	.242 / .390	.241 / .382	.241 / .363
Unmodulated	Ventral	.216 / .274	.216 / .340	.198 / .301	.186 / .294	.204 / .302
ReGraph	Dorsal	.243 / –	.242 / .366	.245 / .390	.244 / .351	.244 / .369
ReGraph	Ventral	.216 / .276	.216 / .340	.193 / .313	.178 / .283	.201 / .303

Appendix

Table 13: Mean significance thresholds and gridness scores (Top-5% retention rate). Each cell shows

\bar{\theta}_{\text{null}}\,/\,\bar{g}_{\text{sig}}

; “–” indicates no grid-like bases were identified. Bold highlights ReGraph’s dorsal stream.

Layer	3%	5%	7%
L1	0.0	0.0	0.2
L2	1.4	5.1	5.6
L3	12.1	8.8	8.3
L4	20.2	15.3	12.1

Appendix

Table 14: Ratio of grid-like bases (%) in the ReGraph dorsal stream across retention rates. The monotonic increase across layers is consistently reproduced at all rates.

Figure 5: RPA pipeline. A. a) SSV2 attention matrices ( $\mathcal{M}$ ) decomposed via HOSVD to isolate spatial bases ( $U_{2},U_{3}$ ). b-c) Top 3, 5, 7% of bases (ranked by singular value) reshaped into 2D spatial weighting maps. d) Gridness scores computed per map. B. a) Logistic-regression classifier trained on spatiotemporally pooled representations. b) Evaluation on the test set (unseen objects) assesses context-invariance; smaller accuracy drop indicates higher invariance.

Model	Initialization	Fine-tuning	Top-1
VideoMAE	K400 (self-supervised)	SSV2-33 subset	80.4
ReGraph (ours)	depth/flow/saliency/object		74.6
SlowFast	K400 (supervised)		66.1
TimeSformer	ImageNet-21K $\rightarrow$ K400 (supervised)		62.3

Appendix

Table 15: Action-recognition comparison on the 33-class SSV2 subset. Top-1 accuracy (%) under the shared 6-view evaluation protocol.

Model	L1	L2	L3	L4
ReGraph — dorsal	0.0	5.1	8.8	15.3
ReGraph — ventral	0.1	6.5	6.0	7.2
VideoMAE	2.4	1.6	1.0	0.9
TimeSformer	1.1	0.9	1.0	1.4
SlowFast — fast	1.6	1.6	0.0	0.0
SlowFast — slow	1.6	0.0	0.0	0.0

Appendix

Table 16: Ratio of grid-like bases (%) in baseline models.

Figure 6: Protocols for testing the functional relevance of grid-like bases. A. Bases-ablation Protocol. a) Top 5% spatial bases sorted by gridness into $G_{\mathrm{high}}$ and $G_{\mathrm{low}}$ . b) Attention matrices ablated by projecting out $G_{\mathrm{high}}$ or $G_{\mathrm{low}}$ subspaces. c) OOD accuracy (Acc.) evaluated under Gaussian noise perturbations. d) OOD Acc. gap, $\Delta_{\mathrm{Acc.}}=\mathrm{Acc.}(G_{\mathrm{low}})-\mathrm{Acc.}(G_{\mathrm{high}})$ , quantifies reliance on $G_{\mathrm{high}}$ . B. Connectivity Reconstruction. a) Disjoint top- $k$ ( $G_{\mathrm{high}}$ ) and bottom- $k$ ( $G_{\mathrm{low}}$ ) subsets selected per mode. b) Unseen attention maps reconstructed via bilateral projection. c) Layer-wise attention maps and $R^{2}$ scores evaluated for each group. d) Reconstruction gap, $\Delta R^{2}=R^{2}_{\mathrm{high}}-R^{2}_{\mathrm{low}}$ , assesses preservation of unseen connectivity.

Layer	Number of Ablated Bases ( $k$ )
Layer	$k{=}1$ : $G_{\text{high}}$ / $G_{\text{low}}$ $[\Delta_{Acc.}\!\uparrow]$	$k{=}5$ : $G_{\text{high}}$ / $G_{\text{low}}$ $[\Delta_{Acc.}\!\uparrow]$	$k{=}9$ : $G_{\text{high}}$ / $G_{\text{low}}$ $[\Delta_{Acc.}\!\uparrow]$
$L_{1}$	82.0 / 81.9 [ $-$ 0.1]	81.9 / 81.6 [ $-$ 0.3]	82.0 / 81.7 [ $-$ 0.3]
$L_{2}$	81.3 / 81.5 [ $+$ 0.2]	80.7 / 81.9 [ $+$ 1.1]	82.6 / 82.9 [ $+$ 0.3]
$L_{3}$	76.0 / 78.3 [ $+$ 2.4]	69.7 / 77.5 [ $+$ 7.8]	70.3 / 77.3 [ $+$ 7.0]
$L_{4}$	97.2 / 97.4 [ $+$ 0.2]	93.6 / 96.6 [ $+$ 3.0]	89.1 / 95.1 [ $+$ 6.0]

Appendix

Table 17: OOD performance to the number of ablated bases (

k

). Cells report OOD accuracy under

G_{\mathrm{high}}

G_{\mathrm{low}}

ablation [gap

\Delta_{Acc.}\!\uparrow

]. A larger positive

\Delta_{Acc.}

indicates greater reliance on

G_{\mathrm{high}}

. Bold highlights the prominent reliance surge at higher layers (

L_{3}

–

L_{4}

), which is most pronounced at

k=9

Pool	$k$	$L_{1}$	$L_{2}$	$L_{3}$	$L_{4}$
30%	5	.005 / .000 [ $+$ 0.005]	.001 / .003 [ $-$ .002]	.055 / .000 [ $+$ .055 ]	.105 / .022 [ $+$ .084 ]
30%	9	.025 / .000 [ $+$ 0.025]	.003 / .066 [ $-$ .062]	.120 / .001 [ $+$ .118 ]	.153 / .052 [ $+$ .101 ]
40%	5	.004 / .000 [ $+$ 0.004]	.001 / .000 [ $+$ 0.001]	.048 / .000 [ $+$ .048 ]	.084 / .016 [ $+$ .068 ]
40%	9	.014 / .000 [ $+$ 0.014]	.001 / .008 [ $-$ .007]	.082 / .000 [ $+$ 0.082]	.131 / .037 [ $+$ 0.094]
50%	5	.004 / .000 [ $+$ 0.004]	.000 / .000 [ $+$ 0.000]	.044 / .000 [ $+$ .044 ]	.060 / .009 [ $+$ .051 ]
50%	9	.012 / .000 [ $+$ 0.012]	.001 / .000 [ $+$ 0.001]	.058 / .000 [ $+$ .058 ]	.123 / .022 [ $+$ .101 ]

Appendix

Table 18: Reconstruction

R^{2}

across pool sizes and subset sizes. Each cell reports reconstruction scores under

G_{\mathrm{high}}

G_{\mathrm{low}}

[gap

\Delta_{R^{2}}=R^{2}_{G_{\mathrm{high}}}-R^{2}_{G_{\mathrm{low}}}\uparrow

], averaged over

13\times 8

(class, head) pairs (

k{=}1

omitted as

R^{2}<0.01

). The bottom row reports the layer-wise mean of the gaps (

\Delta_{R^{2}}

). Bold highlights the prominent surge at

L_{3}

and peak at

L_{4}

on average across configurations.

Single-stream references
Variant	Top-1 (%)	Epochs	h/epoch	Wall-clock	GPU-hours
Dorsal-Only	71.69	24	1.7	$\sim$ 43 h	$\sim$ 86
Ventral-Only	60.38	15	0.2	$\sim$ 4 h	$\sim$ 8
Dorsal–Ventral biological trait asymmetry
(a) Trait-Symmetric	69.78	22	0.8	$\sim$ 19 h	$\sim$ 38
(b) Dorsal-trait Only	72.42	29	1.3	$\sim$ 38 h	$\sim$ 76

Appendix

Table 19: Compute resources and wall-clock training time for each reported ablation. All models trained on 2

\times

NVIDIA RTX 5090 GPUs (32 GB each), with maximum 100 epochs and early stopping at patience

=

8. Wall-clock includes both training and the final 6-view ensemble evaluation; GPU-hours

=

wall-clock

\times

Explore similar work

Emergent Generalization by Representation Learning in Artificial Neural Networks
Jul 11, 2026Hardik Rajpal, Dan GoodmanNeural Circuits Neural Network
1I-X Centre for AI in Science, Imperial College London, W12 0BZ, UK · Department of Electrical and Electronic Engineering, Imperial College London, SW7 2AZ, UK
Self-organized MT Direction Maps Emerge from Spatiotemporal Contrastive Optimization
May 12, 2026Zhaotian Gu, Molan Li, Jie Su +3Human Visual Cortex Contrastive Learning
School of System Science, Beijing Normal University, Beijing 100875, China · Qiyuan Laboratory, Beijing 100095, China · State Key Laboratory of Cognitive Neuroscience and Learning, Beijing Normal2026 University, Beijing 100875, China
Structure Abstraction and Generalization in a Hippocampal-Entorhinal Inspired World Model
May 15, 2026Tianqiu Zhang, Muyang Lyu, Xiao Liu +1Brain-Inspired Synergistic Framework World Models
Peking-Tsinghua Center for Life Sciences, Academy for Advanced Interdisciplinary Studies, IDG/McGovern Institute for Brain Research, Center of Quantitative Biology, School of Psychological and Cognitive Sciences, Key Laboratory of Machine Perception (Ministry of Education), Peking University. · HHMI Janelia.

Jul 11, 2026, q-bio.NCJ/K move · Enter open · S save

Emergent Generalization by Representation Learning in Artificial Neural Networks

Hardik Rajpal, Dan Goodman

1I-X Centre for AI in Science, Imperial College London, W12 0BZ, UK · Department of Electrical and Electronic Engineering, Imperial College London, SW7 2AZ, UK

Neural Circuits Neural Network Representation Learning Emergence Maze Task

Open paper AskPDF arXiv

Dimensionality reduction has proven powerful for identifying neural manifolds, which are low-dimensional structures underlying high-dimensional neural activity. These low-dimensional representations have improved the interpretability of population-level coding. Yet whether such low-dimensional representations are biologically relevant and confer functional advantages in learning systems, or merely reflect neuron-level activity, remains contested in neuroscience. We show that an explicit information bottleneck forcing a recurrent neural network to learn a low-dimensional representation is necessary for rotational and out-of-distribution generalisation in a time-series prediction task. Using information-theoretic measures of causal emergence, we characterise the dynamics of this representation across the memorisation-to-generalisation transition, finding a non-monotonic trajectory which shows an initial decrease, a minimum, and a subsequent rise to a maximum, even as prediction loss falls monotonically. This trajectory scales with task complexity, and the magnitude of emergent structure reliably predicts generalisation performance. Analysis of CA1 hippocampal activity in mice learning an alternating maze task reveals analogous non-monotonic emergence dynamics that track behavioural performance. Together, these findings indicate that the ability of neural networks to learn compact, distributed and emergent representations confers a functional advantage for generalisation, supporting a causal role for learned representations in cognition.

Abstract

Explore similar work

Emergent Generalization by Representation Learning in Artificial Neural Networks

Self-organized MT Direction Maps Emerge from Spatiotemporal Contrastive Optimization

Structure Abstraction and Generalization in a Hippocampal-Entorhinal Inspired World Model