Pith. sign in

REVIEW 2 major objections 5 minor 53 references

A Quotient Homology Theory of Representation in Neural Networks

T0 review · 2 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read The paper claims that for a ReLU network whose polyhedral pieces intersect the input manifold convexly, the homology of the network's representation is isomorphic to the quotient homology of the input manifold by the overlap…

desk verdict Genuinely new overlap decomposition and a promising quotient-homology approach, but Theorem 3.2 rests on an unproved continuity step in B.5; worth refereeing, needs a fix. read the letter →

arxiv 2502.01360 v4 pith:NZ27TM4B submitted 2025-02-03 cs.LG math.ATq-bio.NC

classification cs.LGmath.ATq-bio.NC MSC 55N1068T07
keywords quotienthomologyoverlapdecompositionReLUneuralnetworkspolyhedralBettinumberstopologicaldataanalysisrepresentationslinearprogramming
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper's thesis is that a ReLU network changes the topology of its input manifold only through gluing, and that this gluing can be read off from the network's polyhedral decomposition. It defines the overlap decomposition $\mathcal{O}_\Phi$: the collection of input points whose network outputs lie in the intersection of the affine images of two or more polyhedra. The central theorem states that whenever each intersection $M \cap G_J$ between the input manifold and a polyhedron is convex, the homology groups of the network's representation are isomorphic to the quotient homology groups $H_k(M/\mathcal{O}_\Phi)$, so Betti numbers can be computed without choosing a metric in the output space. This matters because persistent homology mixes geometric information such as curvature and convexity into its estimates, whereas the paper's quotient computation is purely topological. The paper also supplies a linear-programming and union-find algorithm for computing the overlap decomposition and shows on toy problems that it tracks topology rather than geometry.

What carries the argument

The load-bearing object is the overlap decomposition $\mathcal{O}_\Phi$, defined as the equivalence classes generated by points of different polyhedra whose images under the network intersect; it is computed by solving feasibility linear programs over the H-representations of the polyhedra and then running union-find on the detected overlapping pairs. The identity that carries the argument is $H_k(\Phi(M)) \simeq H_k(M/\mathcal{O}_\Phi)$, proven by showing that $M/\sim_\Phi$ is homotopy equivalent to $M/\mathcal{O}_\Phi$ when every $M \cap G_J$ is convex. This homotopy uses the fact that low-rank equivalence classes inside a convex intersection are themselves convex and hence contractible, so they can be collapsed without altering homology. The same machinery also yields the coarser, algorithmically convenient decomposition $\widehat{\mathcal{O}}_\Phi$ used in the numerical experiments.

What would settle it

Take a circle $S^1$ embedded in $\mathbb{R}^2$ and a ReLU network whose polyhedron intersects $S^1$ in a non-contractible arc and whose affine map on that polyhedron has rank one; if the computed $H_1(\Phi(M))$ differs from $H_1(M/\mathcal{O}_\Phi)$, then the theorem's convexity hypothesis is load-bearing and not merely technical.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is Theorem 3.2: given a ReLU network $\Phi$ with polyhedral decomposition $\{G_J\}$ such that $M \cap G_J$ is convex for every $J$, there is an isomorphism $H_k(\Phi(M)) \simeq H_k(M/\mathcal{O}_\Phi)$. The proof separates all non-injectivity into a rank source (an affine map of low rank inside one polyhedron) and an overlap source (points in different polyhedra mapped to the same output), and shows that under the convex-intersection condition the rank source contributes only contractible equivalence classes, so it can be quotiented out without changing homology. Consequently the Betti numbers of the neural representation are determined by the topology of the input manifold together with the gluing pattern encoded in the overlap decomposition, and no metric on the output space is needed. The paper further shows numerically that this overlap-based computation differs from persistent homology, which flags geometric near-identifications as topological features, and that training decreases the volume of overlap regions while increasing their number.

Load-bearing premise

The load-bearing condition is that every intersection $M \cap G_J$ between the input manifold and a polyhedron of the network is convex (or at least contractible), which the paper supports with a heuristic polyhedron-shrinking argument and an MNIST histogram rather than a proof or a checkable condition.

Editorial extensions

If this is right

  • If Theorem 3.2 is correct, the Betti numbers of a neural representation are intrinsic: they depend only on the input manifold's topology and the network's gluing pattern, not on any metric in the output space.
  • Layer-by-layer topology estimates based on persistent homology (including the reproduced Naitzat et al. curves) partly measure geometric distortion; the paper's quotient homology shows purely topological simplification happens more gradually across layers.
  • For a convex input manifold the convex-intersection condition holds automatically, so the quotient-homology computation applies without further assumptions.
  • Counting overlap regions at initialization offers an expressivity measure for non-injective maps, distinct from counting linear regions, and the number of such regions tends to increase with training while their total volume decreases.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A consequence the paper leaves implicit: because quotient Betti numbers depend only on the activation pattern and input topology, they could serve as a scale-free invariant for comparing architectures, removing the need to choose a persistence scale.
  • The convex-intersection hypothesis could be checked per polyhedron on real datasets by testing whether the populated points in each $M \cap G_J$ are linearly separable; such a test would tell practitioners when the rank source can actually be ignored.
  • If polyhedral representations that scale better with dimension (as the paper suggests) are used, the same overlap-finding pipeline could extend quotient homology to high-dimensional inputs where listing H-representations is infeasible.
  • The quotient construction is close to a Reeb space; requiring $x$ and $y$ to lie in the same connected component of $\Phi^{-1}(\Phi(x))$ would give a continuous counterpart whose homology could be approximated by the discrete overlap decomposition.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. This paper introduces an 'overlap decomposition' of a ReLU network's polyhedral decomposition, defined by nontrivial intersections of the images of the affine maps on different polyhedra. It claims (Theorem 3.2) that, when the intersection of each polyhedron with the input manifold is convex, the homology of the network's output representation is isomorphic to the quotient homology H_k(M/O_Φ) of the input manifold by the overlap decomposition. The paper proposes a linear-programming and union-find algorithm to compute the overlap decomposition from a finite sample, and reports experiments on toy curves, on a reproduction of Naitzat et al.'s layer-wise Betti number study, and on sphere classification, showing that quotient homology tracks topological rather than geometric changes and that overlap regions shrink after training.

Significance. The proposed invariant is genuinely interesting: if Theorem 3.2 holds, it gives an intrinsic, metric-free way to compute homology of neural representations from the input manifold's topology and the network's polyhedral structure, avoiding the geometric contamination of persistent homology. The computational pipeline (LP feasibility plus union-find) is concrete, and the empirical comparisons with persistent homology on toy data are thoughtful. The paper also explicitly identifies limitations (type 1/2 errors, convexity heuristics), which is commendable. However, the central proof currently has a gap in the continuity of the homotopy equivalence, and the convexity hypothesis is not yet justified beyond heuristics.

major comments (2)
  1. [Appendix B, Theorem B.5] The proof defines π and the homotopy F piecewise, but it never establishes continuity of π along the boundary between the rank-only region and the full-rank region, nor does it show that the paths γ_z(t) can be chosen continuously in [x]_p. Even the existence of a single path in M/O_Φ is not immediate: M∩G_J convex implies the rank fiber is convex in M, but the quotient by O_Φ can identify points of that fiber with points elsewhere, so the image of the fiber in M/O_Φ need not be convex or even contractible. The paper needs either an explicit continuous selection/deformation retraction or an appeal to a cell-like map theorem, together with a proof that the quotient map is cell-like under the stated hypotheses.
  2. [Section 3.2 and end of Appendix B] The hypothesis 'M∩G_J is convex (or contractible) for every G_J' is load-bearing for Theorem 3.2, yet the paper justifies it by a heuristic scaling argument and the empirical histogram in Figure 6. For non-convex input manifolds (e.g., the spheres in Section 4.3), the condition can fail, and the paper falls back on the observed rarity of low-rank regions (bottom-right of Figure 2) without proving that the rank source is then homology-trivial. The paper itself states that the prevalence of the convexity condition is 'more of a heuristic argument.' Please prove a sufficient condition for the rank source to be homology-invariant when the convexity assumption fails, or provide a testable condition under which Theorem 3.2 applies.
minor comments (5)
  1. [Abstract and Section 3.1] The claim that the overlap decomposition is 'exactly determined' is too strong, because Algorithm 1 only considers populated polyhedra and Section 5.1 acknowledges type 2 errors from unpopulated polyhedra; please qualify the exactness statement (e.g., 'exact with respect to the sampled polyhedra').
  2. [Definition 2.3, Eq. (4)] The set-builder notation for ∼_Φ is circular because it uses [x] inside the definition of [x]; define it instead as the quotient set of M by the equivalence relation x∼y iff Φ(x)=Φ(y).
  3. [Eqs. (5) and (7)] The phrase 'for any I⊊K' is ambiguous because K is not introduced; it should read 'for every index set K with I⊊K'.
  4. [Appendix C.1, Algorithm 1] The loops 'for ∀y∈P_i and ∀z∈P_j' range over continuous sets; the pseudocode should specify that the loops iterate over the data points contained in each polyhedron.
  5. [Figure 2, bottom-right panel] The text says low-rank regions 'appear very rarely,' but no quantitative threshold or fraction is reported; please provide the proportion of low-rank polyhedra among all populated polyhedra.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the quotient-homology isomorphism is a derived conditional statement, and the overlap decomposition is not fitted to the Betti numbers it is used to compute. The paper's self-citation to its own rank-decomposition work is not load-bearing, although Theorem B.5 has a continuity gap that is a correctness risk rather than a circular one.

full rationale

I walked the derivation chain. Theorem 3.2 postulates convexity of M∩G_J and asserts H_k(Φ(M)) ≃ H_k(M/O_Φ). This does not reduce to the theorem's inputs by construction: O_Φ is defined from the polyhedral images (Def. 2.3) and computed by linear-programming feasibility (Eq. 6), without using the target homology groups as a fitting target. The proof route is B.1–B.5: Φ(M) ≅ M/∼_Φ is a standard homeomorphism (Thm B.3), and B.5 attempts an independent homotopy equivalence between M/∼_Φ and M/O_Φ. The rank-decomposition idea is attributed to Beshkov & Einevoll (2024), but Theorem B.2 restates the rank/overlap dichotomy and B.5 gives its own convex-fiber argument; the self-citation is therefore not the load-bearing justification. The Naitzat reproduction chooses ε and k on known ground-truth topology, but this is stated openly in Appendix C and affects only the persistent-homology baseline, not the quotient-homology construction. The paper itself flags a real weakness: "At this stage this is more of a heuristic argument" (Appendix B) about how often the convexity condition holds, and B.5's homotopy F requires a continuous choice of representatives and paths that is asserted, not established. Those are rigor concerns, not cases of a derivation being equivalent to its inputs. Hence no circular step is identifiable.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central theorem rests on standard facts about ReLU networks and on the convexity assumption for manifold-polyhedron intersections. The computational method introduces hand-chosen thresholds and a bounding box, and it relies on dataset coverage. No new physical entities are postulated.

free parameters (3)
  • sensitivity threshold δ = 1 (Section 4.3), 10 (Section 4.2)
    Hand-chosen distance threshold in input space to skip LP overlap checks; affects which overlaps are found and runtime.
  • bounding box for polytope package = [-100, 100] per dimension
    Bounding box used to compute H-representations of polyhedra; restricts the domain and can affect which overlaps are detected.
  • persistent homology scale epsilon and k for Naitzat reproduction = epsilon=2.5, k=14 (D-I), k=19 (D-II, D-III)
    Chosen so that the reproduced PH pipeline recovers ground-truth Betti numbers of the input datasets; a fit to known topology, not used in quotient homology.
assumptions (4)
  • standard math ReLU neural networks are exactly continuous piecewise-linear maps (Arora et al. 2018)
    Used throughout to justify the polyhedral decomposition of the input space.
  • domain assumption The input manifold M is compact and the dataset is sampled from it
    Theorem B.1 uses compactness to conclude injective implies homeomorphism; the datasets are toy manifolds.
  • domain assumption For all polyhedra G_J, the intersection M∩G_J is convex (or contractible)
    This is the load-bearing premise of Theorem 3.2; the paper only provides heuristic and empirical support (Figure 6), not a proof.
  • standard math Intersections of images of polyhedra are convex sets
    Used in the homotopy equivalence between the overlap decomposition and the coarser computed decomposition.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Quotient Homology Theory of Representation in Neural Networks." pith.science (2026). https://pith.science/paper/NZ27TM4B

@misc{pith2026250201360,
  author       = {Pith},
  title        = {Pith review of: A Quotient Homology Theory of Representation in Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NZ27TM4B}},
  note         = {Machine review of arXiv:2502.01360}
}
abstract

Previous research has proven that the set of maps implemented by neural networks with a ReLU activation function is identical to the set of piecewise linear continuous maps. Furthermore, such networks induce a hyperplane arrangement splitting the input domain of the network into convex polyhedra $G_J$ over which a network $\Phi$ operates in an affine manner. In this work, we leverage these properties to define an equivalence relation $\sim_\Phi$ on top of an input dataset, which defines a quotient space that can be split into two sets related to the local rank of $\Phi_J$ and the intersections $\cap \text{Im}\Phi_{J_i}$. We refer to the latter as the \textit{overlap decomposition} $\mathcal{O}_\Phi$ and prove that if the intersections between each polyhedron and an input manifold are convex, the homology groups of neural representations are isomorphic to quotient homology groups $H_k(\Phi(\mathcal{M})) \simeq H_k(\mathcal{M}/\mathcal{O}_\Phi)$. This lets us intrinsically calculate the Betti numbers of neural representations without the choice of an external metric. We develop methods to numerically compute the overlap decomposition through linear programming and a union-find algorithm. Using this framework, we perform several experiments on toy datasets showing that, compared to standard persistent homology, our overlap homology-based computation of Betti numbers tracks purely topological rather than geometric features. Finally, we study the evolution of the overlap decomposition during training on several classification problems and discuss some shortcomings of our method.

Figures

Figures reproduced from arXiv: 2502.01360 by the authors.

Figure 1
Figure 1. A) Illustration of the steps of our method. The manifold [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Reproduced Betti numbers from Naitzat et al. (green) and a quotient homology calculation [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. (Top row) Visualization of the polyhedral decomposition at initialization (magenta) and after training (cyan) across layers. In the last layer we see the points in the overlap decomposition and note that some of them overlap with each other both before and after training. (Bottom row) (left) Polyhedron volume decreases across layers and dimensions before (circles) and after (stars) training. (center) Overlap volume … view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Parameter values that generate the same Betti numbers as the ground truth data. The [PITH_FULL_IMAGE:figures/full_fig_p024_4.png]
Figure 5
Figure 5. Figure 5: More examples of non-linear manifolds and their homology groups. The numbers in the [PITH_FULL_IMAGE:figures/full_fig_p025_5.png]
Figure 6
Figure 6. Figure 6: Histograms of the number of points that fall in each polyhedron induced by four layer [PITH_FULL_IMAGE:figures/full_fig_p026_6.png]
Figure 7
Figure 7. Figure 7: (Top left) Time in seconds to compute the H-representations of all polyhedra in a two [PITH_FULL_IMAGE:figures/full_fig_p026_7.png]
Figure 8
Figure 8. Figure 8: Same plot as in Figure 3 but using an orthogonal initialization (in magenta). [PITH_FULL_IMAGE:figures/full_fig_p027_8.png]
Figure 9
Figure 9. Figure 9: (left) Number of populated polyhedra across 10 randomly initialized networks on the two [PITH_FULL_IMAGE:figures/full_fig_p027_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 46 canonical work pages

  1. [1]

    Adams, H., Emerson, T., Kirby, M., Neville, R., Peterson, C., Shipman, P., Chepushtanova, S., Hanson, E., Motta, F., and Ziegelmeier, L. (2017). Persistence images: A stable vector representation of persistent homology. Journal of Machine Learning Research , 18(8):1--35

  2. [2]

    Andreeva, R., Limbeck, K., Rieck, B., and Sarkar, R. (2023). Metric space magnitude and generalisation in neural networks. In Topological, Algebraic and Geometric Learning Workshops 2023 , pages 242--253. PMLR

  3. [3]

    Arora, R., Basu, A., Mianjy, P., and Mukherjee, A. (2018). Understanding deep neural networks with rectified linear units. International Conference on Learning Representations

  4. [4]

    Balestriero, R., Cosentino, R., Aazhang, B., and Baraniuk, R. (2019). The geometry of deep networks: Power diagram subdivision. Advances in Neural Information Processing Systems , 32

  5. [5]

    Bauer, U. (2021). Ripser: efficient computation of V ietoris- R ips persistence barcodes. J. Appl. Comput. Topol. , 5(3):391--423

  6. [6]

    A rank decomposition for the topological classification of neural representations

    Beshkov, K. and Einevoll, G. T. (2024). A rank decomposition for the topological classification of neural representations. arXiv preprint arXiv:2404.19710

  7. [7]

    J., and Simsekli, U

    Birdal, T., Lou, A., Guibas, L. J., and Simsekli, U. (2021). Intrinsic dimension, persistent homology and generalization in neural networks. Advances in Neural Information Processing Systems , 34:6776--6789

  8. [8]

    and Brun, M

    Blaser, N. and Brun, M. (2022). Relative persistent homology. Discrete & Computational Geometry , 68(4):949--963

Show all 53 references
  1. [9]

    Bubenik, P. et al. (2015). Statistical topological data analysis using persistence landscapes. J. Mach. Learn. Res. , 16(1):77--102

  2. [10]

    Bubenik, P., Hull, M., Patel, D., and Whittle, B. (2020). Persistent homology detects curvature. Inverse Problems , 36(2):025008

  3. [11]

    Carlsson, G. (2009). Topology and data. Bulletin of the American Mathematical Society , 46(2):255--308

  4. [12]

    Dijkstra, E. W. (1959). A note on two problems in connexion with graphs. Numerische mathematik , 1(1):269--271

  5. [13]

    Dupuis, B., Deligiannidis, G., and Simsekli, U. (2023). Generalization bounds using data-dependent fractal dimensions. In International Conference on Machine Learning , pages 8922--8968. PMLR

  6. [14]

    Edelsbrunner, H., Harer, J., and Patel, A. K. (2008). Reeb spaces of piecewise linear mappings. In Proceedings of the twenty-fourth annual symposium on Computational geometry , pages 242--250

  7. [15]

    Fan, F.-L., Huang, W., Zhong, X., Ruan, L., Zeng, T., Xiong, H., and Wang, F. (2023). Deep relu networks have surprisingly simple polytopes. arXiv preprint arXiv:2305.09145

  8. [16]

    Gabrielsson, R. B. and Carlsson, G. (2019). Exposition and interpretation of the topology of neural networks. In 2019 18th ieee international conference on machine learning and applications (icmla) , pages 1069--1076. IEEE

  9. [17]

    Grigsby, J. E. and Lindsey, K. (2022). On transversality of bent hyperplane arrangements and the topological expressiveness of relu neural networks. SIAM Journal on Applied Algebra and Geometry , 6(2):216--242

  10. [18]

    Guss, W. H. and Salakhutdinov, R. (2018). On characterizing the capacity of neural networks using algebraic topology. arXiv preprint arXiv:1802.04443

  11. [19]

    Guti \'e rrez-Fandi \ n o, A., P \'e rez-Fern \'a ndez, D., Armengol-Estap \'e , J., and Villegas, M. (2021). Persistent homology captures the generalization of neural networks without a validation set. arXiv preprint arXiv:2106.00012

  12. [20]

    and Rolnick, D

    Hanin, B. and Rolnick, D. (2019). Deep relu networks have surprisingly few activation patterns. Advances in neural information processing systems , 32

  13. [21]

    Hatcher, A. (2002). Algebraic topology . Cambridge University Press

  14. [22]

    He, K., Zhang, X., Ren, S., and Sun, J. (2015). Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision , pages 1026--1034

  15. [23]

    Hornik, K. (1991). Approximation capabilities of multilayer feedforward networks. Neural networks , 4(2):251--257

  16. [24]

    and Hall, J

    Huangfu, Q. and Hall, J. J. (2018). Parallelizing the dual revised simplex method. Mathematical Programming Computation , 10(1):119--142

  17. [25]

    Kleinberg, J. (2006). Algorithm Design . Addison Wesley

  18. [26]

    and Althoff, M

    Kochdumper, N. and Althoff, M. (2019). Representation of polytopes as polynomial zonotopes. arXiv preprint arXiv:1910.07271

  19. [27]

    Kruskal, W. H. and Wallis, W. A. (1952). Use of ranks in one-criterion variance analysis. Journal of the American statistical Association , 47(260):583--621

  20. [28]

    M., Peterson, C., and Kirby, M

    Liu, Y., Cole, C. M., Peterson, C., and Kirby, M. (2023). Relu neural networks, polyhedral decompositions, and persistent homology. In Topological, Algebraic and Geometric Learning Workshops 2023 , pages 455--468. PMLR

  21. [29]

    Masden, M. (2022). Algorithmic determination of the combinatorial structure of the linear regions of relu neural networks. arXiv preprint arXiv:2207.07696

  22. [30]

    M \'e moli, F., Wan, Z., and Wang, Y. (2022). Persistent laplacians: Properties, algorithms and implications. SIAM Journal on Mathematics of Data Science , 4(2):858--884

  23. [31]

    F., Pascanu, R., Cho, K., and Bengio, Y

    Montufar, G. F., Pascanu, R., Cho, K., and Bengio, Y. (2014). On the number of linear regions of deep neural networks. Advances in neural information processing systems , 27

  24. [32]

    Munkres, J. (2000). Topology . Pearson, second edition

  25. [33]

    Munkres, J. R. (2018). Elements of algebraic topology . CRC press

  26. [34]

    Naitzat, G., Zhitnikov, A., and Lim, L.-H. (2020). Topology of deep neural networks. Journal of Machine Learning Research , 21(184):1--40

  27. [35]

    M., Carlsson, G

    Papamarkou, T., Birdal, T., Bronstein, M. M., Carlsson, G. E., Curry, J., Gao, Y., Hajij, M., Kwitt, R., Lio, P., Di Lorenzo, P., et al. (2024). Position: Topological deep learning is the new frontier for relational learning. In Forty-first International Conference on Machine Learning

  28. [36]

    Papyan, V., Han, X., and Donoho, D. L. (2020). Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences , 117(40):24652--24663

  29. [37]

    Pascanu, R., Montufar, G., and Bengio, Y. (2013). On the number of response regions of deep feed forward networks with piece-wise linear activations. arXiv preprint arXiv:1312.6098

  30. [38]

    and Leit \ a o, A

    Petri, G. and Leit \ a o, A. (2020). On the topological expressive power of neural networks. In TDA \ & \ Beyond

  31. [39]

    T., Goldberg, K., and Kragic, D

    Pokorny, F. T., Goldberg, K., and Kragic, D. (2016). Topological trajectory clustering with relative persistent homology. In 2016 IEEE International Conference on Robotics and Automation (ICRA) , pages 16--23. IEEE

  32. [40]

    Poole, B., Lahiri, S., Raghu, M., Sohl-Dickstein, J., and Ganguli, S. (2016). Exponential expressivity in deep neural networks through transient chaos. Advances in neural information processing systems , 29

  33. [41]

    Raghu, M., Poole, B., Kleinberg, J., Ganguli, S., and Sohl-Dickstein, J. (2017). On the expressive power of deep neural networks. In international conference on machine learning , pages 2847--2854. PMLR

  34. [42]

    Rangamani, A., Lindegaard, M., Galanti, T., and Poggio, T. A. (2023). Feature learning in deep classifiers through intermediate neural collapse. In International Conference on Machine Learning , pages 28729--28745. PMLR

  35. [43]

    Rieck, B., Togninalli, M., Bock, C., Moor, M., Horn, M., Gumbsch, T., and Borgwardt, K. (2018). Neural persistence: A complexity measure for deep neural networks using algebraic topology. arXiv preprint arXiv:1812.09764

  36. [44]

    Serra, T., Tjandraatmadja, C., and Ramalingam, S. (2018). Bounding and counting linear regions of deep neural networks. In International conference on machine learning , pages 4558--4566. PMLR

  37. [45]

    and Althoff, M

    Sigl, S. and Althoff, M. (2023). M-representation of polytopes. arXiv preprint arXiv:2303.05173

  38. [46]

    E., et al

    Singh, G., M \'e moli, F., Carlsson, G. E., et al. (2007). Topological methods for the analysis of high dimensional data sets and 3d object recognition. PBG@ Eurographics , 2:091--100

  39. [47]

    Tralie, C., Saul, N., and Bar-On, R. (2018). Ripser.py : A lean persistent homology library for python. The Journal of Open Source Software , 3(29):925

  40. [48]

    F., and Otter, N

    Turkes, R., Montufar, G. F., and Otter, N. (2022). On the effectiveness of persistent homology. Advances in Neural Information Processing Systems , 35:35432--35448

  41. [49]

    Vick, J. W. (2012). Homology theory: an introduction to algebraic topology , volume 145. Springer Science & Business Media

  42. [50]

    Wasserman, L. (2018). Topological data analysis. Annual Review of Statistics and Its Application , 5(1):501--532

  43. [51]

    Wheeler, M., Bouza, J., and Bubenik, P. (2021). Activation landscapes as a topological summary of neural network performance. In 2021 IEEE International Conference on Big Data (Big Data) , pages 3865--3870. IEEE

  44. [52]

    and Wu, D

    Zhang, X. and Wu, D. (2020). Empirical studies on the properties of linear regions in deep neural networks. arXiv preprint arXiv:2001.01072

  45. [53]

    and Carlsson, G

    Zomorodian, A. and Carlsson, G. (2004). Computing persistent homology. In Proceedings of the twentieth annual symposium on Computational geometry , pages 347--356

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.