Pith. sign in

REVIEW 4 major objections 5 minor 14 references

TAGTorch aims to be the single library for math-aware deep learning tools.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 00:27 UTC pith:RNWPVE64

load-bearing objection A sensible early-stage software announcement whose central claim about a working library needs repository verification; fix the GUI inconsistency and abstract overstatement before publication. the 4 major comments →

arxiv 2607.28755 v1 pith:RNWPVE64 submitted 2026-07-30 cs.LG

TAGTorch: A PyTorch Library for Geometry, Topology, and Symmetry-Aware Machine Learning

Brendan Kennedy , Tegan Emerson , Gregory Roek , Emilie Purvine , Henry Kvinge This is my paper
classification cs.LG
keywords software libraryequivariancetopological data analysisgeometric deep learningsymmetry-aware machine learningpersistent homologyEuler characteristic transformdeep learning software
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper introduces TAGTorch, a package that brings together methods from geometry, topology, and algebra in one deep-learning library. Its goal is to fix a fragmented software ecosystem where mathematical tools for neural networks live in separate, often unmaintained research repositories. TAGTorch currently wraps established equivariant-network libraries and adds native modules for the Euler characteristic transform, persistent homology, and a measure of how equivariant a trained network actually is. The authors argue that a single, extensible interface lowers the barrier for domain scientists and encourages the community to share and compare methods.

Core claim

The central claim is that mathematically inspired methods for machine learning — symmetry-aware architectures, topological data analysis, and model-analysis tools for representation geometry — can be unified in a single deep-learning package without reimplementing mature libraries. The paper describes a design built on group-theoretic abstractions first: protocols for groups, actions, and representations, from which symmetry-based transforms and wrappers are derived. The current release is reported to include wrappers for well-known equivariant-network libraries and native implementations of the Euler characteristic transform, basic persistent homology, and the group empirical equivariance d

What carries the argument

The library's transform-centric API and its group/representation abstraction are the load-bearing pieces. TAGTorch builds from groups to actions to representations, exposing protocols that let external equivariant-network packages plug into a common notation, while transforms serve as the standard way to enter data pipelines. This lets topological operations (such as the Euler characteristic transform and persistence) and symmetry operations (such as data augmentation by group action) be composed like ordinary deep-learning layers.

Load-bearing premise

The load-bearing premise is that the released code actually exists, installs without trouble, and implements the advertised modules and wrappers as described, since the paper provides no installation instructions, API examples, or test reports.

What would settle it

Check the public repository and run the advertised workflows end-to-end: compute the Euler characteristic transform on a small point cloud, build a persistence diagram, and invoke the equivariant wrappers inside a standard training loop. If any advertised function is missing, crashes, or returns results inconsistent with known values, the central claim of a unified, functioning library is falsified.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A researcher can chain a topological feature extractor to an equivariant network in one workflow, instead of stitching separate packages.
  • The group-to-representation abstraction gives a common vocabulary for comparing methods that already share the same underlying mathematics but different code syntax.
  • Because the package is open source and object-oriented, community members can add new methods with minimal changes to existing class hierarchies.
  • If the wrappers behave as described, TAGTorch inherits the reliability of mature equivariant-network implementations while adding missing glue.
  • The interactive persistence-homology widget gives nonexperts a concrete way to see how density and edge-distance hyperparameters change the resulting complex.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Left implicit: the library's success likely hinges less on new algorithms than on steady maintenance and community adoption; a centralized TAG-tools project could become a de facto standard, but it may also overlap with existing topological-deep-learning ecosystems.
  • The transform-centric design could naturally extend to other mathematical pipelines, such as group-equivariant normalization or representation-geometry diagnostics, making it a potential backbone for mathematically structured ML workflows.
  • One testable extension: use the group empirical equivariance deviation to audit off-the-shelf pretrained models, letting users decide whether a symmetry-preserving architecture is worth the extra cost — an application the paper only gestures at.
  • If the library matures to include its planned serialized model recipes, it could lower the entry barrier for domain scientists enough that papers using math-aware layers become reproducible with a few lines of configuration.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces TAGTorch, an open-source PyTorch library intended to unify software for topology-, algebra-, and geometry-inspired machine learning. It motivates the library by the fragmentation of the existing ecosystem and describes design goals: centralizing existing tools, implementing missing methods, lowering the barrier to entry, and ensuring extensibility. The current capabilities are stated to include wrappers for e3nn and escnn, the Euler characteristic transform, basic persistent homology, and one model-analysis function, the group empirical equivariance deviation. The paper also describes planned visualization and interactive GUI features, model configuration via YAML/plaintext, and training workflows. No code listing, installation instructions, version/commit identifier, test report, or API example is provided; the only pointer to the software is a GitHub URL.

Significance. If the described library exists and is maintained, TAGTorch addresses a real and well-documented need: the fragmentation of geometric, topological, and symmetry-aware machine learning software. The design goals are sensible and the paper gives a useful roadmap, with an appropriate discussion of related work. The explicit statement in §5.3 that model analysis is currently limited to a single function is an honest limitation. However, the paper's central claim is that a working software release exists, and the manuscript provides no independently verifiable artifact beyond a URL. The internal inconsistencies between sections about which features are in the initial release further weaken the claim. The significance is therefore conditional: the contribution is a promising project description rather than a demonstrated, verifiable library release.

major comments (4)
  1. [Abstract and §5.3, §A.0.2] The abstract claims the library includes 'training techniques, and model analysis tools.' §5.3 states that model-properties modules 'will include' many functions and that 'At present we have a function' for group empirical equivariance deviation only. Similarly, §A.0.2 says training workflows 'will integrate' with existing libraries, indicating future intent. The abstract's wording overstates the current release. Please align the claims of current versus planned functionality.
  2. [§4.2, Appendix A.0.1, §6] There is an internal inconsistency about the interactive GUI. §4.2 says a visualization pipeline 'is currently in development,' Appendix A.0.1 says 'the initial release of TAGTorch features an interactive Graphical User Interface,' and §6 lists 'interactive Graph User Interfaces (GUI)' under 'Active areas of development.' These statements cannot all describe the same release. Clarify which GUI components, if any, are actually present in the initial release.
  3. [§1 (GitHub link) and general] The central claim is that TAGTorch is a functioning library, but the paper supplies only the URL https://github.com/hkvinge/tagtorch. There is no commit hash, release version, installation instructions, test report, or API example. Without any reproducible artifact, a reader cannot distinguish a working library from a stub. Since software releases are the core of this submission, include a versioned/archived reference (e.g., Zenodo DOI), a minimal installation/usage example for each claimed module, and evidence of smoke tests.
  4. [§5.1, §5.2] The paper states that e3nn/escnn wrappers, the Euler characteristic transform, and 'basic persistent homology capabilities' are currently included, but it does not describe their interfaces, dependencies, or data formats. The claim 'Currently, the library includes topological transforms and data symmetries' in §4.1 is also vague. Please provide concrete API signatures or a code listing for at least these claimed native modules, so reviewers can assess whether the implementations match the surrounding mathematical claims.
minor comments (5)
  1. [Abstract] Typo: 'introducingTopology' is missing a space. Also 'TAGTorchand' and 'ofthe' in the abstract have spacing issues.
  2. [Figure 1 caption] The caption says 'A diagram of the current structure of TAGTorch,' but several boxes (e.g., tagtorch.flows, tagtorch.plot) are not described as implemented anywhere in the text. Either label the diagram as an architectural target or distinguish implemented components from planned ones.
  3. [§4.2] The phrase 'parameter choice in persistence homology' should be 'persistent homology' (the correct term is used in §A.0.1).
  4. [§6] 'Graph User Interfaces' appears to be a typo for 'Graphical User Interfaces' in the list of active development areas.
  5. [§5.2] The list of planned methods (Morse skeletonization, Mapper, vectorizations, zig-zag persistence, sheaf-theoretic methods) would be clearer if a table or explicit marker distinguished 'currently available' from 'planned' entries.

Circularity Check

0 steps flagged

No circularity: TAGTorch is a software-presentation paper with no derivation chain; the sole self-citation is a reimplemented prior method, not load-bearing evidence.

full rationale

This paper makes no mathematical derivation or prediction that could reduce to its inputs. Its central claim is that the TAGTorch library exists and provides certain modules: wrappers for e3nn and escnn, the Euler characteristic transform, persistent homology, and a group empirical equivariance deviation function. No equation is derived from an assumed ansatz, no fitted parameter is relabeled as a prediction, and no uniqueness theorem is invoked. The only self-citation, Kvinge et al. (2022), is the source of a metric that TAGTorch re-implements as one module (Section 5.3); the paper does not use that citation as evidence that the metric is correct, and the metric itself is externally published and independently checkable. None of the enumerated circularity patterns apply. A non-circularity concern exists: the paper asserts a functioning library but gives only an unversioned GitHub URL with no commit hash, installation instructions, tests, or API examples, and it is internally inconsistent about the GUI (Section 4.2 says it is 'currently in development', Appendix A.0.1 says 'the initial release of TAGTorch features an interactive Graphical User Interface', and Section 6 lists interactive GUIs under 'Active areas of development'). These are verifiability and internal-consistency issues, not circular reasoning, so the circularity score remains 0.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

This is a software paper, so the ledger is about the code functioning as described and the underlying math being correctly relayed through wrapper libraries. There are no fitted parameters or invented entities.

axioms (3)
  • ad hoc to paper The linked GitHub repository exists, installs, and contains the described modules.
    The entire central claim (that TAGTorch is an available library) depends on the repository at github.com/hkvinge/tagtorch; no commit hash or installation instructions are provided in the text.
  • domain assumption The wrapped libraries (e3nn, escnn) and their mathematical behaviors remain stable and correctly encapsulated behind TAGTorch's API.
    TAGTorch's stated strategy is to wrap mature libraries rather than re-implement; if the wrappers leak syntax/format differences, the claimed unification fails.
  • standard math Standard mathematical results from representation theory and persistent homology are correctly encoded in the native modules.
    Sections 5.1-5.2 present ECT and persistence as included; correctness is not demonstrated in the paper.

pith-pipeline@v1.3.0-alltime-deepseek · 8722 in / 11721 out tokens · 128049 ms · 2026-08-03T00:27:54.845822+00:00 · methodology

0 comments
read the original abstract

Over the last decade, neural networks have been applied to an increasingly diverse range of applications, including data with rich geometric, topological, or symmetry-related structure. As a result, researchers have increasingly drawn inspiration from topology, algebra, and geometry. Despite this rich algorithmic development, the supporting software ecosystem remains fragmented. Many important methods exist only as research prototypes in unmaintained repositories. We address this by introducing Topology, Algebra, and Geometry Torch (TAGTorch), an open-source, PyTorch-based library that unifies tools inspired by topology, algebra, and geometry, including data-preprocessing methods, architectures, training techniques, and model analysis tools. We describe the design philosophy of TAGTorch and then discuss its current architecture and capabilities, highlighting areas where it can fill gaps in the current software ecosystem. We conclude with a discussion of our future development priorities for the library.

Figures

Figures reproduced from arXiv: 2607.28755 by Brendan Kennedy, Emilie Purvine, Gregory Roek, Henry Kvinge, Tegan Emerson.

Figure 1
Figure 1. Figure 1: A diagram of the current structure of TAGTorch. This will be a starting point upon which we will build out specific capabilities. geometry, model symmetries, loss landscape geometry, and mathematical interpretability techniques. We particularly hope to implement missing links and engineer high-level flows to reduce the barrier to entry into the world of TAG and deep learning. Our strategy blends an appreci… view at source ↗
Figure 2
Figure 2. Figure 2: TAGTorch enables Human-in-the-Loop (HIL) workflows that are necessary to tune the process of topological data transforms. As it matures, TAGTorch will support model specification through serialized configs (e.g., JSON or YAML files), defining model sizes, layer types, and preprocessing. Doing so will allow internal and community contributors to add configurations that define “recipes” for building networks… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

14 extracted references · 1 canonical work pages

  1. [1]

    Git re-basin: Merging models modulo permutation symmetries.arXiv preprint arXiv:2209.04836,

    Samuel K Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries.arXiv preprint arXiv:2209.04836,

    Pith/arXiv arXiv

  2. [8]

    Geomstats: A python package for riemannian geometry in machine learning.Journal of Machine Learning Research, 21 (223):1–9, 2020a

    Nina Miolane, Nicolas Guigui, Alice Le Brigant, Johan Mathe, Benjamin Hou, Yann Thanwerdas, Stefan Heyder, Olivier Peltre, Niklas Koep, Hadi Zaatiti, Hatem Hajri, Yann Cabanes, Thomas Gerald, Paul Chauchat, Christian Shewmake, Daniel Brooks, Bernhard Kainz, Claire Donnat, Susan Holmes, and Xavier Pennec. Geomstats: A python package for riemannian geometry...

    2020

  3. [9]

    Theodore Papamarkou, Tolga Birdal, Michael Bronstein, Gunnar Carlsson, Justin Curry, Yue Gao, Mustafa Hajij, Roland Kwitt, Pietro Li` o, Paolo Di Lorenzo, et al

    URLhttps://github.com/NVIDIA/ cuEquivariance. Theodore Papamarkou, Tolga Birdal, Michael Bronstein, Gunnar Carlsson, Justin Curry, Yue Gao, Mustafa Hajij, Roland Kwitt, Pietro Li` o, Paolo Di Lorenzo, et al. Posi- tion paper: Challenges and opportunities in topological deep learning.arXiv preprint arXiv:2402.08871, 2,

    Pith/arXiv arXiv

  4. [11]

    Frame averaging for invariant and equivariant network design.arXiv preprint arXiv:2110.03336,

    Omri Puny, Matan Atzmon, Heli Ben-Hamu, Ishan Misra, Aditya Grover, Edward J Smith, and Yaron Lipman. Frame averaging for invariant and equivariant network design.arXiv preprint arXiv:2110.03336,

    Pith/arXiv arXiv

  5. [14]

    Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Russ R Salakhut- dinov, and Alexander J Smola

    URLhttps: //arxiv.org/abs/1911.08251. Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Russ R Salakhut- dinov, and Alexander J Smola. Deep sets.Advances in neural information processing systems, 30,

    Pith/arXiv arXiv 1911

  6. [2007]

    Julien Tierny, Guillaume Favelier, Joshua A Levine, Charles Gueunet, and Michael Michaux

    doi: 10.2312/SPBG/SPBG07/ 091-100. Julien Tierny, Guillaume Favelier, Joshua A Levine, Charles Gueunet, and Michael Michaux. The topology toolkit.IEEE transactions on visualization and computer graph- ics, 24(1):832–842,

    work page doi:10.2312/spbg/spbg07/

  7. [2010]

    Gabriele Cesa, Leon Lang, and Maurice Weiler

    doi: 10.1007/s10208-010-9066-0. Gabriele Cesa, Leon Lang, and Maurice Weiler. A program to build E(N)-equivariant steerable CNNs. InInternational Conference on Learning Representations,

    doi:10.1007/s10208-010-9066-0

  8. [2013]

    e3nn: Euclidean neural networks.arXiv preprint arXiv:2207.09453,

    9 Kennedy Emerson Roek Purvine Kvinge Mario Geiger and Tess Smidt. e3nn: Euclidean neural networks.arXiv preprint arXiv:2207.09453,

    Pith/arXiv arXiv

  9. [2014]

    Toy models of superposition.arXiv preprint arXiv:2209.10652,

    Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, et al. Toy models of superposition.arXiv preprint arXiv:2209.10652,

    Pith/arXiv arXiv

  10. [2019]

    Jean-Pierre Serre et al.Linear representations of finite groups, volume

    URLhttps://doi.org/10.5281/zenodo.2533369. Jean-Pierre Serre et al.Linear representations of finite groups, volume

    doi:10.5281/zenodo.2533369

  11. [2021]

    Nnsight and ndif: Democratizing access to open-weight foundation model internals

    Jaden Fiotto-Kaufman, Alexander Loftus, Eric Todd, Jannik Brinkmann, Koyena Pal, Dmitrii Troitskii, Michael Ripa, Adam Belfki, Can Rager, Caden Juang, et al. Nnsight and ndif: Democratizing access to open-weight foundation model internals. InInterna- tional Conference on Learning Representations, volume 2025, pages 92337–92370,

    2025

  12. [2022]

    Fast graph representation learning with pytorch geo- metric.arXiv preprint arXiv:1903.02428,

    Matthias Fey and Jan Eric Lenssen. Fast graph representation learning with pytorch geo- metric.arXiv preprint arXiv:1903.02428,

    Pith/arXiv arXiv 1903

  13. [2024]

    The linear representation hypothesis and the geometry of large language models.arXiv preprint arXiv:2311.03658,

    Kiho Park, Yo Joong Choe, and Victor Veitch. The linear representation hypothesis and the geometry of large language models.arXiv preprint arXiv:2311.03658,

    Pith/arXiv arXiv

  14. [2025]

    doi: 10.21105/joss.07243

    ISSN 2475-9066. doi: 10.21105/joss.07243. Max Klabunde, Tobias Schumacher, Markus Strohmaier, and Florian Lemmerich. Similarity of neural network models: A survey of functional and representational measures.ACM Computing Surveys, 57(9):1–52,

    doi:10.21105/joss.07243