acceptodds
Under review as a conference paper at ICLR 2027

CvB: Conditioned Vector-Bundle Autoencoders with an Active Connection

Abstract

Standard autoencoders learn decodable latent spaces, but do not explicitly distin- guish manifold location from local variation or specify how local features should be compared across locations. We introduce Conditioned Vector Bundle Networks (CvB), an autoencoder that organizes latent representations into base coordinates and base-conditioned fibers linked by learned linear transport. The model jointly optimizes reconstruction, fiber alignment, and transport composition consistency, with regularization that discourages base–fiber redundancy and fiber collapse. This construction provides an explicit mechanism for coordinating local features across the learned base. Across ten random seeds, CvB achieves mean adjusted Rand indices of 0.511 on MNIST and 0.405 on Fashion-MNIST, compared with 0.422 and 0.352 for autoencoders with matched latent dimensions. Controlled ablations identify fiber alignment as the principal contributor to these clustering gains. In a separate experiment using 4,000 MNIST training samples, CvB re- duces measured relative topological distortion from 140.58 to 26.28 compared with the matched autoencoder. Digit-label probes further show that identity is concentrated in the base: balanced prediction accuracy reaches 0.741 from base coordinates and 0.196 from fiber coordinates, whereas removing transport shifts identity information into the fiber. These results support learned fiber transport as an inductive bias for organizing decodable representations of manifold-structured data

Primary area:
learning theory
Keywords:
unsupervised representation learning, geometric deep learning, autoencoders, vector bundles, manifold learning, learned transport, fiber alignment, latent disentanglement

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_XshcOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors present an autoencoder with a latent space structured as a vector bundle rather than the ordinary flat vector space used in traditional approaches. The architecture is inspired by two theorems. The authors claim that the proposed approach is the best parametric approach for clustering and topological preservation in image benchmarks and is competitive in interpolation. Moreover, the authors highlight several limitations, including that the proposed approach does not outperform bundle-specialized baselines on raw geodesic distortion; its decoded interpolations trade ambient smoothness for manifold fidelity; its connection is flat by construction; and its disentanglement evidence uses MNIST-derived factors alongside a dSprites factorization.

Critical strengths

The idea and its application look interesting, with potential contributions to the field. However, at this stage, the work is very difficult to understand, and there is some evidence of potential plagiarism.

Critical weaknesses

The presentation is bad; there is no structure, and the text does not smoothly explain the limitations of the literature, with duplicate information appearing in the introduction and related work. The authors present results in the introduction based on their proposed concepts without introducing the concepts first. All of this makes it difficult for the reader to follow. They use non-scientific terminology (“The honest summary is”, “destroys the base/fiber split.”). The authors do not properly reference tables in the appendix; for example, Table 4 is in the appendix. The presentation is not structured; it is chaotic and difficult to follow. Mathematical proofs and explanations need to be distinguished from the experimental approach.

Two theorems are presented by the authors. No reference citations for the foundations of the theorems are included. The way they are presented does not justify whether these are the authors' own theorems or whether they follow from the literature; this needs to be clarified. A literature reference for the starting point is needed if these are their own theorems, and if they are from the literature, appropriate references are required, as at this point it looks like plagiarism. For example, Theorem 3.1 has a basis in: [1] Ambrose, W., & Singer, I. M. (1953). A theorem on holonomy. Transactions of the American Mathematical Society, 75(3), 428–443, which establishes the classical relationship between curvature and holonomy. The authors did not mention anything related to this; [2] Kineider, C., Kydonakis, G., Rogozinnikov, E., Tatitscheff, V., & Thomas, A. (2026). Bundles and Connections. In Spectral Networks, Lecture Notes in Mathematics, Vol. 2386, pp. 45–61. Springer of, Theorem 4.29: Flat bundles, locally constant transition functions and fundamental-group representations, of Theorem 4.30: Riemann–Hilbert correspondence. (plagiarism) Similarly, the same concerns apply to the other theorem.

Regarding the results, the authors mention that UMAP and PaCMAP will be compared and benchmarked against the proposed methodology. However, later in the manuscript, they do not include the results of these techniques, with the explanation: “Non-parametric UMAP/PaCMAP reach higher silhouettes on FashionMNIST (Table 8), but provide no decoder and therefore cannot perform interpolation or semantic transfer.” This explanation is biased, as this metric is one of those the authors use in their validation protocol to define superiority. However, in Table 2, the authors present PCA, even though it is not decodable, and do not present UMAP and PaCMAP. This looks biased again. To that end, this appears to be an unjustified attempt to convince the reader of the superiority of their approach. The authors could have used a fixed decoder for all the interpolation approaches.

At the first level of evaluation, the proposed approach fails to outperform the baselines, and the authors state: “The connection therefore does not claim to improve geodesic preservation directly; its value appears on the interpolation and semantic-clustering tasks that the bundle geometry is designed for. Table 2 summarizes the image benchmarks.” I am not convinced that the technique outperforms the baselines. Their technique appears to achieve higher results on the two metrics that cannot be used by approaches without a decoder or transport mechanism; however, there does not appear to be a significant difference in either the standard deviation or the mean value. No statistical significance analysis is included elsewhere to verify these differences.

Questions for authors

Can the authors clarify whether the two presented theorems are their own original contributions? Could you please justify why you chose this specific validation protocol?

13 comments
AuthorsOpenReview ↗

Response letter

We thank you for the review. The response follows your four comments on weaknesses and two questions in order. The updated table and page numbers refer to the revised draft.

Q1. Are Theorems 3.1 and 3.2 your own original contributions, and is the structure clear?

Response. We sincerely apologize for accidentally missing the references. Although this should not be considered an excuse, it was a strange and unintended oversight.

Theorems 3.1 and 3.2 in Section 3 ("Background: Data Manifolds as Vector Bundles") are not our contribution.** Our proposal/work starts in Section 4** ("The Conditioned Vector Bundle Network"). We use the theorems as the starting point and as the criteria that define and motivate the CvB (Conditioned Vector-Bundle Autoencoder) architecture. We sincerely hope you could reconsider our work, we listed our contribution as below. Thank you very much.

We revised Section 3 (Background section as below). In addition to the references you mentioned, we included further references for Theorems 3.1 and 3.2 within the theorem statements.

Added in main draft-(Section 3 Background; page 3-4; copied part) Both statements are classical and are included for completeness, not as novel theoretical contributions. Theorem 3.1 is the standard equivalence between flatness, path-independent parallel transport, vanishing holonomy, and local triviality of a vector bundle; it is the content of the Ambrose--Singer holonomy theorem \cite{ambrose1953theorem} together with the standard transition-function formulation of connections \cite{kobayashi1996foundations,lee2018introduction,Kineider2026} (see in particular Theorems 4.29--4.30 of \cite{Kineider2026} on flat bundles, locally constant transition functions, and the Riemann--Hilbert correspondence). Theorem3.2 is the elementary statement that a parallel section of a trivial bundle with the trivial connection is locally constant. We give short proofs in Appendices H and I because we use both statements as operational criteria; the paper's contribution is the architecture that instantiates them and the experiments that test it, not the theorems themselves. We refer the reader to the cited texts for the full theory and do not extend it here.

Added in main draft-(Section 3 Background; page 4; copied part) The following classical fact from the standard theory of connections and transition functions \cite{lee2018introduction, kobayashi1996foundations} underpins our diagnostics. For the transition-function and cocycle formulation see \cite{steenrod1999topology, husemoller1966fibre}; for the equivalent flatness, holonomy, and path-independence formulations see \cite{docarmo1992riemannian, petersen2016riemannian, joyce2007riemannian}.

Added in main draft-(Section 3 Background; page 4; copied part) The elementary fact that parallel sections of a trivial bundle with the trivial connection are locally constant is standard \cite{kobayashi1996foundations,lee2018introduction,docarmo1992riemannian,petersen2016riemannian,joyce2007riemannian}. We show that in this identity-transport limit such a representation cannot organize base-local information in the fiber through the alignment objective. We stress the scope of the statement up front:

The main contribution of our work is the CvB (Conditioned Vector-Bundle Autoencoder) architecture, below we made a comparison table to show the novelty of CvB. The table is added in page 35 as Table 25.

CapabilityCvB (ours)FibeRedChart-AEAE / VAEUMAP / PaCMAP
Decoder (reconstruction, decoded interpolation)✓✗✓✓✗
Learned transport T(bi,bj)T(b_i,b_j)✓✗✗✗✗
Base/fiber split with probes✓✗✗✗✗
Identity-preserving fiber transfer + composition✓n/an/an/an/a
Ground-truth parallel-transport recovery (Appendix D)✓ 7.1∘7.1^\circ/5.5∘5.5^\circn/an/an/an/a
MGDE, MNIST latents (lower better)0.3700.2050.8870.323 (AE)—
RTD, MNIST latents (lower better)26.850.3103.6141.4 (AE)—
Clustering ARI, MNIST0.511 (best with decoder)0.3890.3480.422 (AE)0.667 (no decoder)
Reconstruction MSE, test0.0398n/a0.01510.0216 (AE)n/a

Our contributions are: (1) We propose CvB, an autoencoder whose latent space is a base chart bb for manifold location, a conditioned fiber ff for local variation, and a learned transport map that connects fibers (the connection). The contribution is the network, not the theory: the flatness equivalence and the identity-transport limit are existing classical results (Theorems 3.1 and 3.2, stated with references in Section 3); we use them as design criteria and inspiration, and we test the value of base-conditioning under an active connection directly. (2) An active connection. We show the learned transport is non-trivial and has a measurable effect: under identical architectures, removing the alignment objective reduces transport consistency (from ∼10−6\sim 10^{-6} to O(1)\mathcal{O}(1)) and removes the base/fiber separation. (3) A decodable bundle latent. Unlike non-parametric embeddings or geometry-specialized baselines, CvB provides a single parametric autoencoder for clustering, decoded interpolation, and a probed base/fiber split, at a constant $\sim$3$\times$ per-epoch overhead over a matched flat autoencoder. (4) Empirical evaluation. On the image benchmarks we test, CvB attains the highest clustering ARI among the parametric methods we compare and preserves the intrinsic topology of MNIST latents substantially better than matched flat autoencoders; on ambient interpolation it is competitive but not the best, and we report that tradeoff explicitly.

We have also included the following references, in addition to the references you mentioned, and appended them to Section 3 (Background), where they are related to the theorem.

Cogliati, A. (2016). Schouten, Levi-Civita and the notion of parallelism in Riemannian geometry. Historia Mathematica, 43(4), 427–443. do Carmo, M. P. (1993). Riemannian geometry (F. Flaherty, Trans.). Birkhäuser.

Husemöller, D. (1966). Fibre bundles (Vol. 5). Springer.

Joyce, D. D. (2007). Riemannian holonomy groups and calibrated geometry (Vol. 12). Oxford University Press.

Marle, C.-M. (2007). The works of Charles Ehresmann on connections: From Cartan connections to connections on fibre bundles. Banach Center Publications, 76, 65–86.

Petersen, P. (2016). Riemannian geometry (3rd ed., Vol. 171). Springer.

Steenrod, N. E. (1999). The topology of fibre bundles (Vol. 14). Princeton University Press.

Weatherburn, C. E. (1928). On Levi-Civita's theory of parallelism. Bulletin of the American Mathematical Society, 34(5), 585–590.

Please check. Paragraph in Section 3 (Background) with blue mark; pages 3-4. page 35 Table 25.

Q2. Presentation, structure, and terminology (CW1)

Response. Thank you for your comment, we have restructured the manuscript, and revised the introduction to make the motivation and contribution clear.

-Added in main draft-(Section 1 Introduction; page 2; copied part): Paper organization: Section 2 positions CvB relative to flat, topological, non-parametric, and bundle-structured methods. Section 3 introduces the bundle objects and states the two classical facts used later. Section 4 describes the proposed architecture and objective. Section 5 reports the empirical evaluation; the geometric verification of the learned bundle is in Appendix \B. Mathematical statements and background are confined to Section 3 and Appendices H-I; the remaining appendices contain full experimental tables, figures, and diagnostics, which are cited as 'Appendix Table' or ``Appendix Figure'' in the main text.

-Revised paragraph -(Section 1 Introduction; page 1; copied part) Many tasks that use a learned representation---interpolating between inputs, transferring the style of one object onto the identity of another, and sampling new examples---require its latent space to encode two things: where an example lies among its neighbors, and how it can vary without leaving the data manifold. Autoencoders \cite{pinaya2020autoencoders} and variational autoencoders \cite{kingma2013vae} give smooth, decodable latents, but they treat the latent space as a single flat vector space. In this space every dimension is a global coordinate, so the model can neither separate where an example sits on the manifold from how it varies locally, nor describe how a local feature changes as it moves between regions. The cost is measurable: on image data, a matched flat autoencoder distorts the intrinsic topology of its own latents about 5$\times$ more than our model, and the interpolants it decodes cut straight through ambient space instead of staying on the manifold surface.

-Revised paragraph -(Section 1 Introduction; page 1; copied part) On manifold-structured data \cite{bronstein2017geometric}, this limitation is inherent: the quality of a representation depends on geometry, and a flat latent space cannot encode it. We need a representation with three properties: it must capture geometry, be decodable by a parametric decoder, and support learned transport between regions. Prior work offers at most two of the three. Non-parametric embeddings \cite{mcinnes2018umap,wang2020pacmap} capture geometry but have no encoder or decoder; chart atlases \cite{chen2022chart,schonsheck2022semi} give a parametric autoencoder but no transport between charts; and FibeRed \cite{scoccola2022fibered} provides bundle geometry but is non-parametric, with no decoder or learned transport. Section 2 compares these approaches in detail. CvB is designed to provide all three together.

In addition, the informal phrasing has been removed; we have revised the duplicated passages part, experiment reorganization;

Please check. Paper organization paragraph Sec. 1 page 2; Sec. 3 Background pages 3-4; Appendices H–I pages 31-32.

Reviewer_R47COpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes a new autoencoder that jointly optimizes reconstruction, fiber alignment, transport composition consistency, and discourages base-fiber redundancy and fiber collapse.

Experiments were done to validate the class-discriminative and topology preserving properties of the proposed method. However, it would be more helpful if reconstruction, transport composition, and other properties of a fiber-transport autoencoder could be reported.

Critical strengths

The most conclusive experiments include:

  1. fiber alignment being an effective regularizer for learning more class-discriminative latent representations (Table 2), and
  2. digit-label probe shows that digit identity is concentrated in base (Table 14).

It also learns good topology preserving compared to vanilla autoencoders (Table 11), which is beneficial for interpolations (Table 10).

Critical weaknesses

On the other hand, additional metrics I think are informative to check but not yet included in the paper are:

  1. reconstruction (recovery) error, which is reported in other autoencoder based work (eg, Chart autoencoders Schonsheck et al. (2020)).
  2. fiber transfer effectiveness; maybe this can be measured by attribute transfer accuracy or target identity preservation on some datasets with such annotations?
  3. semantic meaning of the fiber: while Table 14 shows that digit identity is concentrated in base, Table 15 seems to show that recoverability of some style factors seem modest, which may suggest that there might not be an easy way for attribute manipulation using this learned representation suppose one wants to utilize the learned fibers.
  4. transport compositionality: while the model itself is trained with cocycle regularization, it would be more helpful if there is some qualitative illustration of semantic attribute transfer composition.
Questions for authors

Questions raised in above section

4 comments
AuthorsOpenReview ↗

Response letter

We thank you for the review. The response follows your four points in order; the three points about the fiber and transport are answered together because they use the same experiments. All numbers are 10-seed mean ± std unless stated. The updated table and page numbers refer to the revised draft.

The revision adds Section G.5 (reconstruction error, pages 26-27), Section G.6 (fiber transfer and transport composition, pages 27-28), the new paragraphs of Section G.4 (continuous style factors and the dSprites ground-truth study, pages 25), and the scope statement in Section K.7 in page 35.

Q1. Reconstruction error

Response. Thank you for the comment, we added Section G.5 (pages 26-27) and Table 19 in page 27, the experiment is across five capacity-matched autoencoders on MNIST: ntrain=5000n_{\text{train}}=5000, neval=3000n_{\text{eval}}=3000 held out, 50 epochs, batch 256256, Adam 3×10−43\times10^{-4}, 10 seeds. Test MSE:

MethodTest MSE ↓
Chart-AE0.0151 ± 0.0001
AE0.0216 ± 0.0006
CvB (no transport)0.0285 ± 0.0013
VAE0.0344 ± 0.0008
CvB (full)0.0398 ± 0.0030

CvB has the highest reconstruction error of the five. The no-transport control removes 28%28\% of CvB's error and closes 62%62\% of the gap to the plain AE, so the reconstruction cost is attributable to the bundle objective (transport, base/fiber split, alignment), not to capacity or overfitting. Test MSE is above train MSE for every method with a ratio of means between 1.001.00 and 1.151.15.

Note: These numbers are not comparable to the reconstruction column of the scalability study in Section G.9 in page 31, which uses a different encoder/decoder and schedule. The advantages of CvB are summarized in Table 25 of page 35 and below:

CapabilityCvB (ours)FibeRedChart-AEAE / VAEUMAP / PaCMAP
Decoder (reconstruction, decoded interpolation)✓✗✓✓✗
Learned transport T(bi,bj)T(b_i,b_j)✓✗✗✗✗
Base/fiber split with probes✓✗✗✗✗
Identity-preserving fiber transfer + composition✓n/an/an/an/a
Ground-truth parallel-transport recovery (Appendix D)✓ 7.1∘7.1^\circ/5.5∘5.5^\circn/an/an/an/a
MGDE, MNIST latents (lower better)0.3700.2050.8870.323 (AE)—
RTD, MNIST latents (lower better)26.850.3103.6141.4 (AE)—
Clustering ARI, MNIST0.511 (best with decoder)0.3890.3480.422 (AE)0.667 (no decoder)
Reconstruction MSE, test0.0398n/a0.01510.0216 (AE)n/a

Please check. Section G.5 (pages 26-27), Table 19 in page 27, and Table 25 of page 35.

Q2. What does the fiber carry, and can it transfer and compose?

Response. Thank you for your comment, we added three experiments in Section G.4 (Disentanglement: Base-Fiber Separation, page 25) and Section G.6 (Fiber transfer and transport composition, page 27), with the full protocol and controls.

1. Semantic content of the fiber (Section G.4, page 25). Block probes on MNIST show that identity is concentrated in the base: base alone 0.743±0.0210.743 \pm 0.021, fiber alone 0.188±0.0530.188 \pm 0.053, full code 0.7440.744. The no-transport ablation reverses the split: base 0.523±0.0450.523 \pm 0.045, fiber 0.808±0.0220.808 \pm 0.022, full code 0.8530.853. The unconditioned-but-transported variant matches full CvB (base 0.7430.743, fiber 0.1870.187), so the separation is produced by the active transport, not by conditioning the fiber entry. Continuous factors are linearly recoverable from the fiber at R2=0.375R^2=0.375 (intensity) and 0.3710.371 (width) on MNIST. On dSprites with ground-truth factors, the fiber carries scale R2=0.366R^2=0.366 and position R2=0.806R^2=0.806; the discrete shape factor is not concentrated in the base (0.0140.014) because dSprites has no class/identity structure. MIG is 0.0280.028 for CvB versus 0.0950.095 for β\beta-TCVAE and 0.0490.049 for FactorVAE, the expected ordering on a dataset designed for total-correlation objectives (Tables 16, 17, 18).

Table 16: MNIST: disentanglement metrics (10 seeds). Standard axis-alignment scores are uniformly low because pixel-derived factors are not axis-aligned in any latent. The informative signal is the block-level digit accuracy: CvB concentrates identity in the base (base 0.743 ≈ full 0.744, fiber 0.188), while the connection-absent ablation places identity information in the fiber (fiber 0.808).

MethodBase-Acc ↑Fiber-AccFull-Acc ↑β-cons. ↑DCI total ↑SAP ↑
CvB (ours)0.743 ± 0.0210.188 ± 0.0530.744 ± 0.0200.422 ± 0.0710.0858 ± 0.00150.006
CvB (uncond. fiber)0.743 ± 0.0230.187 ± 0.0600.744 ± 0.0230.461 ± 0.0620.0856 ± 0.00220.006
CvB (no transport)0.523 ± 0.0450.808 ± 0.0220.853 ± 0.0060.633 ± 0.0840.0868 ± 0.00310.005
AE----0.883 ± 0.0030.636 ± 0.0480.0876 ± 0.00150.005
VAE----0.732 ± 0.0110.428 ± 0.0510.0857 ± 0.00160.004
β-VAE----0.451 ± 0.0340.269 ± 0.0240.0840 ± 0.00110.004

Table 17: MNIST: linear recoverability (R², ridge, 5-fold CV) of continuous pixel factors from the fiber block. Higher means the factor is better retained in the fiber. Columns: stroke intensity, stroke width, vertical position. Total-correlation baselines (FactorVAE [kim2018factorvae], β-TCVAE [chen2018betatc]) are evaluated on the same protocol in Table 13; a standardized factorized dSprites benchmark is reported in Table 18.

MethodR² intensity ↑R² width ↑R² vpos ↑
CvB (ours)0.375 ± 0.1090.371 ± 0.107-0.001 ± 0.002
CvB (uncond. fiber)0.552 ± 0.1250.546 ± 0.1240.003 ± 0.004
CvB (no transport)0.863 ± 0.1020.854 ± 0.1020.013 ± 0.009

Table 18: dSprites: MIG (mean over the five ground-truth factors, higher is better) and linear R² (ridge, 5-fold CV). "Base"/"Fiber" split the code into CvB's base and fiber blocks (flat baselines have no such split; --- marks not-applicable); "R² full" is from the whole code. "vpos" is the mean of the horizontal and vertical position factors. 10 seeds. In the table, "CvB (no tran.)" denotes CvB (no transport).

MethodMIG ↑R² full ↑shape R² base ↑shape R² fiber ↑scale R² base ↑scale R² fiber ↑
CvB (ours)0.028 ± .010.485 ± .010.014 ± .010.027 ± .010.213 ± .090.366 ± .08
CvB (no tran.)0.031 ± .020.559 ± .010.004 ± .010.065 ± .010.130 ± .170.853 ± .03
FactorVAE0.049 ± .030.512 ± .02------------
β-TCVAE0.095 ± .040.496 ± .02------------
β-VAE0.062 ± .030.491 ± .02------------
VAE0.035 ± .020.543 ± .01------------
AE0.026 ± .010.546 ± .01------------
Methodvpos R² base ↑vpos R² fiber ↑
CvB (ours)0.912 ± .030.806 ± .07
CvB (no tran.)0.675 ± .320.335 ± .27

2. Transfer effectiveness (Section G.6 page 27). Protocol: 200 held-out (source, donor) triples on MNIST per seed; encode, swap the fiber block, decode along the direct route TkiT_{ki} and the composed route TjiTkjT_{ji}T_{kj}. Identity is read by a 5-nearest-neighbor digit probe fitted on real pixels (probe accuracy 0.936±0.0030.936 \pm 0.003 on real held-out images; chance 0.10.1). Attribute adoption uses intensity, stroke width, and vertical position with the gap-closed gain. We report two anti-collapse regimes, because the paper default λvar=10−2\lambda_{\mathrm{var}}=10^{-2} reduces the fiber amplitude and makes the swap numerically negligible:

Metric (10 seeds)(a) default λvar=10−2\lambda_{\mathrm{var}}{=}10^{-2}(b) fiber-alive λvar=10−1\lambda_{\mathrm{var}}{=}10^{-1}
CvB (full): identity recon → direct0.644 → 0.6440.613 → 0.615
CvB (full): attribute gain−0.000 ± 0.001+0.003 ± 0.005
CvB (no transport): identity recon → direct0.823 → 0.1040.847 → 0.063
CvB (no transport): attribute gain+0.292 ± 0.111+0.319 ± 0.127
CvB (full): cocycle residual ↓0.155 ± 0.0280.190 ± 0.030
CvB (no cocycle): cocycle residual ↓0.905 ± 0.0660.839 ± 0.081

Full CvB preserves source identity through both routes in both regimes, including regime (b) where the swap changes the pixels. The no-transport ablation drops identity to chance (0.823→0.1040.823\to0.104, 0.847→0.0630.847\to0.063) and adopts donor style (+0.292+0.292/+0.319+0.319), which is the failure mode the base/fiber split is designed to avoid. Attribute adoption for full CvB is ≈0\approx 0 in both regimes, so the metric detects the difference between identity-preserving transport and style transfer. The current objective supports identity-preserving transport, not style manipulation;

3. Transport compositionality (Section G.6 pages 27-28, Figure 8 in page 29). The figure shows source | donor | recon | direct | composed | no-transport for held-out MNIST triples; the direct and composed routes agree, and the no-transport column adopts the donor. The quantitative companions are: relative cocycle residual 0.155±0.0280.155 \pm 0.028 versus 0.905±0.0660.905 \pm 0.066 without the cocycle term; MSE(direct, composed) 1.9×10−101.9\times10^{-10} versus 2.4×10−92.4\times10^{-9}; direct and composed identity means agree to within 0.0020.002 in every condition (largest per-seed gap 0.0100.010). On held-out base triples the absolute residual is 0.029±0.0130.029 \pm 0.013 versus 1.25±0.121.25 \pm 0.12, and the 4-hop error is 0.021±0.0120.021 \pm 0.012 versus 0.779±0.0650.779 \pm 0.065 (Appendix B-C in pages 13-14, Table 3 in page 15, Figure 1 in page 14).

Table 3: Held-out transport diagnostics, mean ± std over 10 seeds. Lipschitz: finite-difference slope of ||T(bᵢ + εu, bⱼ) - T(bᵢ, bⱼ)||_F in ε. Cocycle: ||TⱼₖTᵢⱼ - Tᵢₖ||_F on held-out triples. Hop-4: relative error of the 4-hop composition vs. the direct transport. Gauge: ||Tᵢⱼ - gⱼgᵢ⁻¹||_F / ||Tᵢⱼ||_F with gᵢ = T(h → i). σ_min and κ (median): singular values of T on held-out edges; sing.%: fraction with σ_min < 10⁻³. Fiber-ratio: neighbor-pair fiber variation normalized by total fiber variance.

VariantLipschitz ↓Cocycle ↓Hop-4 ↓Fiber-ratio ↓σ_min ↑
CvB (full)0.004 ± 0.0020.03 ± 0.010.02 ± 0.0120.049 ± 0.0082.9×10⁻⁴ ± 1.1×10⁻⁴
CvB (no cocycle)0.15 ± 0.0141.25 ± 0.120.78 ± 0.0650.047 ± 0.0130.175 ± 0.058
CvB (no transport)0000.0853 ± 0.00721
CvB (T=I + align)0000.0482 ± 0.00291
VariantSing.% ↓Gauge ↓ConnErr ↓κ ↓
CvB (full)96.5 ± 3.1715 ± 20791.7×10⁻⁵ ± 1.8×10⁻⁵2.2×10⁴ ± 4.9×10⁴
CvB (no cocycle)0.36 ± 0.313.86 ± 2.593.8×10⁻⁵ ± 4.7×10⁻⁵8.8 ± 4.7
CvB (no transport)003.82 ± 0.861
CvB (T=I + align)009.0×10⁻⁷ ± 1.2×10⁻⁷1

Please check. Section G.4 (page 25) and Section G.6 (page 27), as well as Figure 1 (page 14) and Figure 8 (page 29).

Closing

Thank you for your thoughtful review and encouraging comments. We have updated the manuscript with additional experimental results and a revised Introduction, and we invite you to review the updated version.

Sincerely, The Authors

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

The authors adopt a differential-geometric approach to designing an autoencoder that aims to learn a manifold-structured latent representation using vector bundles, where each point on a base manifold is associated with a vector space called a fiber. Specifically, the autoencoder's latent space is a base chart that encodes the position and a fiber that encodes the local variation, and a learned linear transport map that connects the fibers. The authors state two theorems (3.1 and 3.2) that lay the basis for their framework. Their network architectre is trained on multiple losses (reconstruction, fiber aligment, cocyle consisteny, an orthogonal penalty (between the base and the fiber), and a variance constraint on the fiber. The authors motivate the autoencode architecture based on geometric concepts such as connections and parallel transport. They present results on datasets such as MNIST, meshes (in R^3), FashionMNIST, etc. and show improvements for some conditions. While the ideas raised in the paper are interesting, there are several weaknesses related to how the theory connects with the implementation, as well as the experimental results.

Critical strengths

The motivation for this work is interesting. While there are several appraoches for latent representation learning and analysis, the idea of imposing a geometric structure that induces vector bundle learning along with constraints on the fibers and their base points. To my knowledge the only other work that compares to this is FibeRed (Scoccola and Perea), which the authors cite.

The authors perform extensive empirical evaluation on several datasets and other baseline methods. That's a strength. The method performs well on clustering criteria over other methods, especially on MNIST.

The ablation studies are fairly performed. The authors are upfront about negative results (effect of cocyle loss).

Critical weaknesses

There was a disconnect between the differential geometric theory and the implementation in terms of the neural network architecture. The motivation for having such a connection is clear. However, the authors didn't satisfactorily bridge this gap. They outlined several theorems and theoretical results, however, they did not adeuquately show whether the results, learned representations satisfied the mathematical properties of the objects. The smoothness properties of hte manifold, the parallel transport operation etc. are not clear. Thus even if the authors show a learned linear transport, how can one ensure that it applies to some manifold with a vector bundle.

The losses, including the the alignment loss, cocycle loss, orthogonal loss, fiber variance, can be thought to have numerical properties and could also be interpreted as constraints on matrices and linear operators without having a geometric connection. Thus the proposed architecture may be recovering some still meaningful latent features and the transformations between them, without the need of having the geometric machinry.

The theroms 3.1 and 3.2 are established results. However, the authors don't verify whether the proposed neural architecture satisfies these properties. The authors should clarify what is mathematically novel about these results and how they fit into the proposed neural network architecture.

A vector bundle is a global geometric structure realized as a set of locally defined vector spaces. The authors motivate their method by claiming to learn the transport map (pairwise map), however, it was not clear how they fit with the global bundle structure. Likewise, the base positions, fiber features (local variation) while learnt and generated by the neural network, were disconnected with the local manifold geometry.

The authors mention that CvB is different from FibeRed in that it offers an autoencoder architecture that estimate the base coodinates, fibers and pairwise maps (transport). However, as stated above, it was not clear whether the learned fibers and transport operators retain the geometric properties motivating the original vector-bundle formulation. Further in FibeRed, a global embedding is obtained from local geometric representations. But here it seems CvB is learning the fiber features and pairwise transformations through reconstruction and alignment losses, which could be defined without geometric constraints. Additionally, in the experimental results, FibeRed is generally performing much better (if not the best) without the auto-encoder architecture. The FibeRed comparison was absent from the MNIST topological latent representations in Figure 11.

The authors mention that the cocycle loss did not contribute to the change (improvement or degradation) in performance in the experimental results. That itself is not a limitation, as it may be enforcing some other geometric properties. Thus, minimizing the cocycle loss (|| T_{jk} T_{ij} - T_{ik}|) may leaad to the compositional consistency as the authors state. However, this is not demonstrated anywhere. Thus even if the term did not cause any meaningful change in learning the model, the authors could through empirical results show what it means to enforce this consistency.

Since the paper makes strong claims about modeling the vector bundle and fiber properties of the underlying data, one could simplify the experimental data with known bundle structure, fibers etc. MAybe consider a special tangent space representation on spherical manifolds and test whether their learned transformations can learn the geometric structure. This may be a direct test of the framework, as to whether geometry is learnt or whether just feature alignment is taking place.

Questions for authors

How can the authors better verify the mathematical properties that are learned through their architecture? While a vector-bundle-inspired latent architecture need not explicitly estimate tangent spaces, the authors should clarify what geometric properties characterize the learned fibers and how the local transformations establish the claimed global structure. Without such evidence, the model appears closer to a structured latent representation with learned pairwise feature transformations than to a demonstrated recovery of vector-bundle geometry.

The authors should provide more discussion and explanation on how one can connect the main results of theorems 3.1 and 3.2 to the neural network architecture, and importantly the evaluation of results.

There is another concern for the cocycle loss. Invertibility is not guaranteed anywhere, thus it may produce degenerate solutions (although it doesn't appear to be the case from the experimental results). Can the authors verify this?

Is there a better synthetic dataset or manifolds with explicit known structures, that could be simulated so the authors could demonstrate that the learned transport recovers accurate local variation or recovers known parallel transport results?

8 comments
AuthorsOpenReview ↗

Response letter

We thank you for the review. Below, we summarize your questions and comments. Points addressed by the same experiment are grouped together. Unless otherwise stated, all reported values are the mean ± standard deviation over 10 random seeds. The updated table and page numbers refer to the revised draft.

Q1. Theorems 3.1 and 3.2, how do they connect to the network and its evaluation?

Response. The two theorems are classical results, and we have added the relevant references (we are sorry, we missed important references). They serve as the foundation and defining criteria for the CvB architecture. We also added a paragraph in Section 3 clarifying that these results are included for completeness rather than as novel theoretical contributions. The paper's main contribution is the CvB architecture and its evaluation. To connect the theorems to CvB, we conducted additional experiments (Appendices B and C), whose results are summarized under Q2-Q7 below.


Q2. Are the learned fibers and transport operators the mathematical objects?

Response. We added Appendix B, "Verifying the geometric properties of the learned bundle", and Appendix C with the full numbers. Each property is measured on held-out data against three controls: no cocycle, no transport (T≡IdT\equiv\mathrm{Id}), and identity-transport plus alignment. The experiments are:

  1. Smoothness. We perturb one base argument of T(bi,bj)T(b_i,b_j) by εu\varepsilon u and measure the finite-difference Lipschitz slope on held-out pairs. The slope is 0.0042±0.00190.0042 \pm 0.0019 for CvB versus 0.149±0.0140.149 \pm 0.014 without the cocycle term, and the response grows continuously with ε\varepsilon (Table 3, page 15). The learned transport is therefore a continuous, locally Lipschitz function of its base arguments, consistent with a smooth connection.

Table 3: Held-out transport diagnostics, mean ± std over 10 seeds. Lipschitz: finite-difference slope of ||T(bᵢ + εu, bⱼ) - T(bᵢ, bⱼ)||_F in ε. Cocycle: ||TⱼₖTᵢⱼ - Tᵢₖ||_F on held-out triples. Hop-4: relative error of the 4-hop composition vs. the direct transport. Gauge: ||Tᵢⱼ - gⱼgᵢ⁻¹||_F / ||Tᵢⱼ||_F with gᵢ = T(h → i). σ_min and κ (median): singular values of T on held-out edges; sing.%: fraction with σ_min < 10⁻³. Fiber-ratio: neighbor-pair fiber variation normalized by total fiber variance.

VariantLipschitz ↓Cocycle ↓Hop-4 ↓Fiber-ratio ↓σ_min ↑
CvB (full)0.004 ± 0.0020.03 ± 0.010.02 ± 0.0120.049 ± 0.0082.9×10⁻⁴ ± 1.1×10⁻⁴
CvB (no cocycle)0.15 ± 0.0141.25 ± 0.120.78 ± 0.0650.047 ± 0.0130.175 ± 0.058
CvB (no transport)0000.0853 ± 0.00721
CvB (T=I + align)0000.0482 ± 0.00291
VariantSing.% ↓Gauge ↓ConnErr ↓κ ↓
CvB (full)96.5 ± 3.1715 ± 20791.7×10⁻⁵ ± 1.8×10⁻⁵2.2×10⁴ ± 4.9×10⁴
CvB (no cocycle)0.36 ± 0.313.86 ± 2.593.8×10⁻⁵ ± 4.7×10⁻⁵8.8 ± 4.7
CvB (no transport)003.82 ± 0.861
CvB (T=I + align)009.0×10⁻⁷ ± 1.2×10⁻⁷1
  1. Transport is active. The held-out connection error is ∼10−6\sim 10^{-6} for CvB versus O(1)\mathcal{O}(1) for every connection-absent ablation, five orders of magnitude on identical architectures (Table 5 (page 18), Table 3). The transport is also non-identity: ∥T−Id∥F=3.24±0.06\|\mathcal{T}-\mathrm{Id}\|_F = 3.24 \pm 0.06 (Table 1, page 8).

Table 5: Connection error 𝔼||T(bᵢ, bⱼ)fᵢ - fⱼ||² (lower = transport better aligns fibers). The connection error is near-zero for CvB and CvB (align. only).

MethodMNISTFashionMNIST
CvB (ours)7.30×10⁻⁶ ± 2.30×10⁻⁶8.71×10⁻⁶ ± 3.02×10⁻⁶
CvB (no transport)3.29 ± 0.191.35 ± 0.10
CvB (orth. only)3.78 ± 0.282.76 ± 0.13
CvB (align. only)8.88×10⁻⁶ ± 2.15×10⁻⁶6.44×10⁻⁶ ± 7.86×10⁻⁷
CvB (cocyc. only)2.82 ± 0.122.87 ± 0.15

Table 1: Spectral diagnostics of the learned transport T(bᵢ, bⱼ) over held-out base pairs. σ_min is the mean smallest singular value; det T > 0 is the fraction of orientation-preserving operators. The transport is full-rank and non-identity; the regularized variant improves conditioning.

| Method | median cond | mean σ_min | det T > 0 | ||𝒯 - Id||_F | | :--- | :---: | :---: | :---: | :---: | | CvB (ours) | 8823 ± 7084 | 4.4×10⁻⁴ ± 0.7×10⁻⁴ | 0.49 ± 0.24 | 3.24 ± 0.06 | | CvB (uncond. fiber) | 9546 ± 8036 | 4.2×10⁻⁴ ± 1.1×10⁻⁴ | 0.28 ± 0.21 | 3.38 ± 0.09 | | CvB (conditioned T) | 6244 ± 2753 | 5.2×10⁻⁴ ± 1.1×10⁻⁴ | 0.51 ± 0.22 | 2.97 ± 0.07 | | CvB (no transport) | 1.00 | 1.00 | 1.00 | 0.00 |

  1. Composition and path independence. On held-out base triples the cocycle residual is ∥TjkTij−Tik∥F=0.029±0.013\|T_{jk}T_{ij}-T_{ik}\|_F = 0.029 \pm 0.013 with the term versus 1.25±0.121.25 \pm 0.12 without it; the 4-hop composition error is 0.021±0.0120.021 \pm 0.012 versus 0.779±0.0650.779 \pm 0.065 (Appendix B (pages 13-14), Figure 1 (page 14), Table 3).

  2. Local-to-global structure. Fix a hub hh, define gi:=T(h→i)g_i := T(h\to i), and measure ∥Tij−gjgi−1∥F/∥Tij∥F\lVert T_{ij}-g_jg_i^{-1}\rVert_F/\lVert T_{ij}\rVert_F on held-out edges. The mean is 715±2079715 \pm 2079 and the median is 18.618.6 (Appendix B, Appendix C, pages 13–14). The residual is conditioning-limited: TT has σmin⁡∼3×10−4\sigma_{\min}\sim 3\times 10^{-4}, and the single seed with κ=1.7×105\kappa = 1.7\times 10^{5} contributes 6.9×1036.9\times 10^{3} of the mean. We therefore claim local bundle consistency on MNIST, and we test global trivialization where it is decidable: on the flat cylinder, the learned map recovers the ground-truth parallel transport to 7.12±5.197.12 \pm 5.19 (Appendix D, pages 14–15).

  3. Base/fiber versus local geometry. The base block reads digit identity at 0.743±0.0210.743 \pm 0.021 versus 0.188±0.0530.188 \pm 0.053 from the fiber alone (full code 0.7440.744); base trustworthiness against the true coordinates is 0.969±0.0120.969 \pm 0.012 on the sphere. Continuous factors are linearly recoverable from the fiber: R2=0.375R^2 = 0.375 (intensity) and 0.3710.371 (width) on MNIST; on dSprites, scale 0.3660.366 and position 0.8060.806 (Section G.4 (page 25), Tables 16, 17, 18 (pages 25-26); Appendix D (pages 14-16)). Table 16: MNIST: disentanglement metrics (10 seeds). Standard axis-alignment scores are uniformly low because pixel-derived factors are not axis-aligned in any latent. The informative signal is the block-level digit accuracy: CvB concentrates identity in the base (base 0.743 ≈ full 0.744, fiber 0.188), while the connection-absent ablation places identity information in the fiber (fiber 0.808).

MethodBase-Acc ↑Fiber-AccFull-Acc ↑β-cons. ↑DCI total ↑SAP ↑
CvB (ours)0.743 ± 0.0210.188 ± 0.0530.744 ± 0.0200.422 ± 0.0710.0858 ± 0.00150.006
CvB (uncond. fiber)0.743 ± 0.0230.187 ± 0.0600.744 ± 0.0230.461 ± 0.0620.0856 ± 0.00220.006
CvB (no transport)0.523 ± 0.0450.808 ± 0.0220.853 ± 0.0060.633 ± 0.0840.0868 ± 0.00310.005
AE----0.883 ± 0.0030.636 ± 0.0480.0876 ± 0.00150.005
VAE----0.732 ± 0.0110.428 ± 0.0510.0857 ± 0.00160.004
β-VAE----0.451 ± 0.0340.269 ± 0.0240.0840 ± 0.00110.004

Table 17: MNIST: linear recoverability (R², ridge, 5-fold CV) of continuous pixel factors from the fiber block. Higher means the factor is better retained in the fiber. Columns: stroke intensity, stroke width, vertical position. Total-correlation baselines (FactorVAE [kim2018factorvae], β-TCVAE [chen2018betatc]) are evaluated on the same protocol in Table 13; a standardized factorized dSprites benchmark is reported in Table 18.

MethodR² intensity ↑R² width ↑R² vpos ↑
CvB (ours)0.375 ± 0.1090.371 ± 0.107-0.001 ± 0.002
CvB (uncond. fiber)0.552 ± 0.1250.546 ± 0.1240.003 ± 0.004
CvB (no transport)0.863 ± 0.1020.854 ± 0.1020.013 ± 0.009

Table 18: dSprites: MIG (mean over the five ground-truth factors, higher is better) and linear R² (ridge, 5-fold CV). "Base"/"Fiber" split the code into CvB's base and fiber blocks (flat baselines have no such split; --- marks not-applicable); "R² full" is from the whole code. "vpos" is the mean of the horizontal and vertical position factors. 10 seeds. In the table, "CvB (no tran.)" denotes CvB (no transport).

MethodMIG ↑R² full ↑shape R² base ↑shape R² fiber ↑scale R² base ↑scale R² fiber ↑
CvB (ours)0.028 ± .010.485 ± .010.014 ± .010.027 ± .010.213 ± .090.366 ± .08
CvB (no tran.)0.031 ± .020.559 ± .010.004 ± .010.065 ± .010.130 ± .170.853 ± .03
FactorVAE0.049 ± .030.512 ± .02------------
β-TCVAE0.095 ± .040.496 ± .02------------
β-VAE0.062 ± .030.491 ± .02------------
VAE0.035 ± .020.543 ± .01------------
AE0.026 ± .010.546 ± .01------------
Methodvpos R² base ↑vpos R² fiber ↑
CvB (ours)0.912 ± .030.806 ± .07
CvB (no tran.)0.675 ± .320.335 ± .27

Please check. Appendix B-C (pages 13-14) and Table 3(page 15); Table 4 (page 16) and Appendix D (pages 14-16); the fiber-probe paragraphs in Section G.4 (pages 25-26); the scope boundaries in Section K.7 (page 35).

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary Response to All Reviewers

Summary Response to All Reviewers

We thank all reviewers for their time, careful reading, and constructive feedback. The comments have directly improved the paper, and we have used them to restructure the manuscript, add new experiments. Below we give a short general response, re-state our contribution, provide the capability comparison, and summarize the point-to-point replies. Detailed point-to-point responses are provided separately for each reviewer (R1, R2, R3).

1. The missing references are our fault, and we have fixed them

We sincerely apologize for accidentally omitting the foundational references for Theorems 3.1 and 3.2 under Background Section 3. This was an unintended oversight and entirely our responsibility. We thank Reviewer 3 for pointing it out and for the specific references.

We have revised Section 3 (Background): Theorems 3.1 and 3.2 are now explicitly stated to be classical background results, not our contribution. They are included for completeness and used as design criteria for the architecture. We added the relevant references.

The paper's contribution begins in Section 4 and is stated next.

2. Our contribution is the CvB architecture

The main contribution of this work is CvB (Conditioned Vector-Bundle Autoencoder): a single parametric autoencoder whose latent space is a vector bundle with three learned components:

  1. A base chart b that encodes where an example lies on the data manifold.
  2. A conditioned fiber f that encodes local variation at that location.
  3. A learned transport map T(b_i, b_j) that connects fibers across base locations, i.e. the connection.

This gives four contributions:

  1. The CvB architecture. We propose an autoencoder whose latent space is explicitly a base chart, a conditioned fiber, and a learned transport map. The contribution is the network, not the classical theory.
  2. An active connection. We show the learned transport is non-trivial and has a measurable effect. Under identical architectures, removing the alignment objective degrades transport consistency from ~1e-6 to O(1) and removes the base/fiber separation.
  3. A decodable bundle latent. Unlike non-parametric embeddings or geometry-specialized baselines, CvB provides one parametric, decodable model that supports clustering, decoded interpolation, and a probed base/fiber split, at a constant ~3x per-epoch overhead over a matched flat autoencoder.
  4. Empirical evaluation. On the image benchmarks we test, CvB attains the highest clustering ARI among parametric methods and preserves the intrinsic topology of MNIST latents substantially better than matched flat autoencoders. On raw geodesic distortion it is competitive but not the best, and we report that tradeoff explicitly.

Novelty

The novelty is not the differential-geometric theory, which is classical, but the architecture that makes it operational in one parametric, decodable model. CvB is, to our knowledge, the first autoencoder that jointly learns a base chart, a conditioned fiber, and an active transport between fibers.

3. What CvB supports compared with the baselines

The table below is added to the revised manuscript as Table 25 (Section K.6). It summarizes the capabilities, not just the scores.

CapabilityCvB (ours)FibeRedChart-AEAE / VAEUMAP / PaCMAP
Decoder (reconstruction, decoded interpolation)✓✗✓✓✗
Learned transport T(b_i, b_j)✓✗✗✗✗
Base/fiber split with probes✓✗✗✗✗
Identity-preserving fiber transfer + composition✓n/an/an/an/a
Ground-truth parallel-transport recovery (Appendix D)✓ 7.1° / 5.5°n/an/an/an/a
MGDE, MNIST latents (lower better)0.3700.2050.8870.323 (AE)—
RTD, MNIST latents (lower better)26.850.3103.6141.4 (AE)—
Clustering ARI, MNIST0.511 (best with decoder)0.3890.3480.422 (AE)0.667 (no decoder)
Reconstruction MSE, test0.0398n/a0.01510.0216 (AE)n/a

We state the tradeoff directly: FibeRed leads raw geodesic fidelity, Chart-AE and AE lead reconstruction, and UMAP/PaCMAP lead raw clustering but have no decoder. CvB does not beat all baselines on every metric. What it provides is a specific combination, a decodable bundle latent with an active, verified connection, at a quantified cost. FibeRed has no decoder and no learned transport, so every transport-bearing task in the paper (decoded interpolation, base/fiber probes, identity-preserving fiber transfer and composition, and recovery of a known parallel transport) is undefined for it. This is why we propose CvB, and we now report the comparison in both directions.

Closing

We thank the reviewers again for the detailed and constructive feedback. It has made the paper clearer and more honest about both its contribution and its limitations. We hope the revised manuscript and the point-to-point responses address the concerns, and we would be grateful if you could check our response and the revised draft.

Sincerely, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.