arxiv: 2605.06218 · v3 · submitted 2026-05-07 · 💻 cs.LG

AffineLens: Capturing the Continuous Piecewise Affine Functions of Neural Networks

Yi Wei , Xuan Qi , Furao Shen , Jian Zhao , Vittorio Murino , Cigdem Beyan This is my paper

Pith reviewed 2026-05-13 07:00 UTC · model grok-4.3

classification 💻 cs.LG

keywords piecewise affine neural networkshyperplane arrangementsaffine region enumerationneural network expressivitypolyhedral partitioncontinuous piecewise affine functionsnetwork geometry

0 comments

The pith

AffineLens enumerates the exact maximal continuous piecewise affine regions of a neural network inside any given bounded input polytope by layer-wise selection of intersecting hyperplanes.

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

The paper introduces AffineLens to compute the precise partition of input space into affine pieces that a neural network creates, instead of using indirect statistics or loose bounds. It takes a calibrated input polytope, discards hyperplanes that miss the domain, and builds the regions sequentially through the layers while guaranteeing each returned region is non-empty and contains an interior point. The method works for standard modern blocks such as convolutions, residuals, batch normalization, and pooling because each preserves the continuous piecewise affine character. Readers care because the tool supplies both visual maps of the regions and quantitative counts, letting them directly measure how architecture choices shape the geometry of the learned function.

Core claim

Given a bounded input polytope, AffineLens identifies the subset of neuron-induced hyperplanes that intersect the domain, enumerates the resulting affine sub-regions in a layer-wise manner, and returns provably non-empty maximal CPA regions together with interior representatives. The framework exploits the fact that fixed activation patterns restrict the network to an affine map, allowing exact enumeration even when the architecture includes batch normalization, pooling, residual connections, multilayer perceptrons, and convolutional layers.

What carries the argument

Layer-wise enumeration of maximal affine regions induced by the subset of neuron hyperplanes that intersect the calibrated input polytope.

If this is right

Networks become directly comparable through region-complexity metrics such as total region count and average region volume.
Decision boundaries and region partitions can be visualized for qualitative inspection of any supported architecture.
Design choices such as depth, width, or skip connections can be evaluated by their effect on the geometry of the induced partition.
Quantitative expressivity studies become feasible without relying on activation histograms or theoretical upper bounds.

Where Pith is reading between the lines

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

The same region enumeration could be used to compute tighter bounds on Lipschitz constants or robustness margins by inspecting the linear maps inside each region.
Controlling region count during training might serve as a new regularizer that limits unnecessary fragmentation of input space.
Safety-critical applications could verify that the learned function satisfies certain geometric properties by inspecting the explicit region list rather than the weights alone.

Load-bearing premise

Every network component, including batch-norm, pooling, residuals and convolutions, preserves the continuous piecewise-affine property so the layer-wise count remains exact.

What would settle it

A concrete counter-example in which the method returns a region that is empty inside the input polytope, or a network component that maps an affine piece to a curved surface.

Figures

Figures reproduced from arXiv: 2605.06218 by Cigdem Beyan, Furao Shen, Jian Zhao, Vittorio Murino, Xuan Qi, Yi Wei.

**Figure 1.** Figure 1: Precise visualization of input distribution, affine regions, and decision boundaries view at source ↗

**Figure 2.** Figure 2: Visualization of the affine region search. (a) Define an input space view at source ↗

**Figure 3.** Figure 3: Two types of data distributions are used in the experiments, each comprising 200 view at source ↗

**Figure 4.** Figure 4: The impact of network depth on the expressivity in terms of the number of rep view at source ↗

**Figure 5.** Figure 5: An analysis of the influence of neurons in shallow and deep layers on the number view at source ↗

**Figure 6.** Figure 6: Arrangement and quantitative analysis of affine regions expressed by MLPs and view at source ↗

**Figure 7.** Figure 7: Arrangement and quantitative analysis of affine regions expressed by MLPs and view at source ↗

**Figure 8.** Figure 8: (a) Dynamic visualization of decision boundary evolution and affine region forma view at source ↗

read the original abstract

Piecewise affine neural networks (PANNs) provide a principled geometric perspective on neural network expressivity by characterizing the input--output map as a continuous piecewise affine (CPA) function whose complexity is governed by the number, arrangement, and shapes of its affine regions. However, existing interpretability and expressivity analyses often rely on indirect proxies (e.g., activation statistics or theoretical upper bounds) and rarely offer practical, accurate tools for enumerating and visualizing the induced region partition under realistic architectures and bounded input domains. In this work, we present AffineLens, a unified framework for computing the hyperplane arrangements and polyhedral structures underlying PANNs. Given a calibrated (bounded) input polytope, AffineLens identifies the subset of neuron-induced hyperplanes that intersect the domain, enumerates the resulting affine sub-regions in a layer-wise manner, and returns provably non-empty maximal CPA regions together with interior representatives. The framework further provides visualizations of region partitioning and decision boundaries, enabling qualitative inspection alongside quantitative region counts. By exploiting the affine restriction property of CPA networks under fixed activation patterns, AffineLens supports a broad class of modern components, including batch normalization, pooling, residual connections, multilayer perceptrons, and convolutional layers. Finally, we use AffineLens to perform a systematic empirical study of architectural expressivity, comparing networks through region complexity metrics and revealing how design choices influence the geometry of learned functions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

AffineLens is a working implementation that enumerates maximal affine regions layer-by-layer for networks with batch-norm, pooling, residuals and convolutions, plus visualizations and counts.

read the letter

The paper's main deliverable is AffineLens, a procedure that takes a trained network and a bounded input polytope, finds the relevant hyperplanes, and returns the maximal non-empty CPA regions with interior points. It does this by propagating activation patterns and polyhedral cells forward through the layers while using linear-programming feasibility checks to drop empty cells. The framework also outputs visualizations of the partition and decision boundaries, and the authors run a small empirical comparison of region counts across architectures. That combination of exact enumeration, modern components, and usable output is what is new here. Earlier region-counting work mostly handled plain ReLU MLPs or gave only upper bounds; this one claims to stay exact on the listed components because each preserves the CPA property. The argument for non-emptiness is standard LP feasibility inside each candidate cell, and the stress-test note confirms no internal contradiction once the construction is written out. The soft spots are practical rather than foundational. The abstract and description give no runtime numbers, no scaling curves, and no direct head-to-head against prior enumeration code, so it is unclear whether the method stays usable beyond toy widths and depths. The empirical section would also be stronger if it included verification on networks whose region counts are already known by hand. This work is aimed at interpretability researchers who want concrete region geometry instead of activation histograms or theoretical bounds. A reader who needs to measure how architecture choices change the number or shape of affine pieces could get immediate value from the tool. I would send it to peer review; the core claims are checkable and the output is a concrete artifact that others can test or build on.

Referee Report

0 major / 3 minor

Summary. The paper introduces AffineLens, a computational framework that takes a calibrated bounded input polytope and performs layer-wise enumeration of the hyperplane arrangements induced by a piecewise-affine neural network (including batch-norm, pooling, residuals, convolutions, and MLPs). It identifies intersecting neuron hyperplanes, enumerates the resulting polyhedral cells, and returns provably non-empty maximal CPA regions together with interior representative points, plus visualizations and quantitative region-complexity metrics for comparing architectural expressivity.

Significance. If the layer-wise propagation and feasibility checks are exact, AffineLens supplies the first practical, architecture-agnostic tool for exact enumeration of maximal affine regions under realistic modern components. This moves beyond theoretical upper bounds or activation statistics and directly supports geometric interpretability, decision-boundary analysis, and controlled empirical studies of how design choices affect function complexity.

minor comments (3)

[§4.3] §4.3: the statement that the LP feasibility check guarantees non-emptiness is correct in principle, but the manuscript should explicitly state the numerical tolerance used and how degenerate (zero-volume) cells are filtered.
[Figure 5] Figure 5: the color scale for region density is not labeled with units or range; readers cannot interpret the quantitative comparison across architectures without it.
[§5.2] The complexity discussion in §5.2 reports empirical runtimes but omits a big-O statement in terms of number of neurons and input dimension; adding this would clarify scalability limits.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive summary of AffineLens and for recommending minor revision. No specific major comments were raised in the report, so we provide no point-by-point responses below. We will incorporate any minor editorial suggestions in the revised version.

Circularity Check

0 steps flagged

No significant circularity; AffineLens is a computational enumeration algorithm

full rationale

The paper describes AffineLens as an algorithmic procedure that, given a bounded input polytope, identifies intersecting neuron hyperplanes, performs layer-wise enumeration of affine sub-regions, and uses standard linear-programming feasibility checks to certify non-empty maximal CPA regions. No equations or steps reduce the output to a quantity defined by the authors' own fitted parameters, self-citations, or ansatzes. The CPA preservation under batch-norm, pooling, residuals, and convolutions follows from well-known properties of these operations (affine or CPA maps), and the 'provably non-empty' guarantee is a direct consequence of polyhedral feasibility rather than any self-referential construction. The contribution is therefore self-contained as a practical tool without circular reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Review performed on abstract only; the ledger is therefore minimal and provisional.

axioms (2)

domain assumption Networks composed of affine layers and piecewise-linear activations (ReLU, etc.) induce continuous piecewise-affine input-output maps.
Standard premise of all PANN literature; invoked implicitly when the framework is said to apply to 'PANNs'.
domain assumption The input domain can be represented as a bounded polytope whose intersection with hyperplanes can be computed exactly.
Required for the 'identifies the subset of neuron-induced hyperplanes that intersect the domain' step.

pith-pipeline@v0.9.0 · 5564 in / 1446 out tokens · 50175 ms · 2026-05-13T07:00:04.653948+00:00 · methodology

Review history (3 revisions) →

discussion (0)

Reference graph

Works this paper leans on

49 extracted references · 49 canonical work pages

[1]

On the Number of Linear Regions of Deep Neural Networks , booktitle =

Guido Mont. On the Number of Linear Regions of Deep Neural Networks , booktitle =

work page
[2]

Kaiming He and Xiangyu Zhang and Shaoqing Ren and Jian Sun , title =. Proc. 2015 , publisher =

work page 2015
[3]

Maas and Awni Y

Andrew L. Maas and Awni Y. Hannun and Andrew Y. Ng , title =. Proc. 2013 , address =

work page 2013
[4]

Advances in Neural Information Processing Systems , volume =

Boris Hanin and David Rolnick , title =. Advances in Neural Information Processing Systems , volume =

work page
[5]

Sergey Ioffe and Christian Szegedy , title =. Proc. Int. Conf. Mach. Learn. (. 2015 , publisher =

work page 2015
[6]

1975 , publisher =

Thomas Zaslavsky , title =. 1975 , publisher =. doi:10.1090/memo/0154 , isbn =

work page doi:10.1090/memo/0154 1975
[7]

2017 , doi =

Alexandru Dimca , title =. 2017 , doi =

work page 2017
[8]

On the Expected Complexity of Maxout Networks , booktitle =

Hanna Tseran and Guido Mont. On the Expected Complexity of Maxout Networks , booktitle =

work page
[9]

Arturs Berzins , title =. Proc. Int. Conf. Mach. Learn. (. 2023 , publisher =

work page 2023
[10]

Baraniuk , title =

Ahmed Imtiaz Humayun and Randall Balestriero and Guha Balakrishnan and Richard G. Baraniuk , title =. Proc. 2023 , publisher =

work page 2023
[11]

Neural Networks , volume =

Hao Chen and Yu Guang Wang and Huan Xiong , title =. Neural Networks , volume =. 2023 , doi =

work page 2023
[12]

Huan Xiong and Lei Huang and Wenston J. T. Zang and Xiantong Zhen and Guo-Sen Xie and Bin Gu and Le Song , title =. 2024 , doi =

work page 2024
[13]

Pawel Piwek and Adam Klukowski and Tianyang Hu , title =. Proc. Conf. Uncertainty in Artificial Intelligence (. 2023 , publisher =

work page 2023
[14]

2025 , doi =

Zhiwei Li and Cheng Wang , title =. 2025 , doi =

work page 2025
[15]

Baraniuk , title =

Randall Balestriero and Richard G. Baraniuk , title =. Proc. Int. Conf. Mach. Learn. (. 2018 , publisher =

work page 2018
[16]

Advances in Neural Information Processing Systems , volume =

Setareh Cohan and Nam Hee Kim and David Rolnick and Michiel van de Panne , title =. Advances in Neural Information Processing Systems , volume =

work page
[17]

Elisenda Grigsby and Kathryn Lindsey , title =

J. Elisenda Grigsby and Kathryn Lindsey , title =. 2022 , doi =

work page 2022
[18]

arXiv preprint arXiv:2401.11188 , year =

Randall Balestriero and Yann LeCun , title =. arXiv preprint arXiv:2401.11188 , year =. doi:10.48550/arXiv.2401.11188 , eprinttype =

work page doi:10.48550/arxiv.2401.11188
[19]

Kaiming He and Xiangyu Zhang and Shaoqing Ren and Jian Sun , title =. Proc. 2016 , publisher =

work page 2016
[20]

Alexis Goujon and Arian Etemadi and Michael Unser , title =. J. Comput. Appl. Math. , volume =. 2024 , doi =

work page 2024
[21]

Lampert , title =

Mary Phuong and Christoph H. Lampert , title =. Proc. Int. Conf. Learn. Represent. (. 2020 , publisher =

work page 2020
[22]

Jeong and David Rolnick , title =

Boris Hanin and Ryan S. Jeong and David Rolnick , title =. Proc. Int. Conf. Learn. Represent. (. 2022 , publisher =

work page 2022
[23]

Yuan Wang , title =. Proc. Int. Joint Conf. Artif. Intell. (. 2022 , publisher =

work page 2022
[24]

Advances in Neural Information Processing Systems , volume =

Saket Tiwari and George Konidaris , title =. Advances in Neural Information Processing Systems , volume =

work page
[25]

Kording , title =

David Rolnick and Konrad P. Kording , title =. Proc. Int. Conf. Mach. Learn. (. 2020 , publisher =

work page 2020
[26]

Xiao Zhang and Dongrui Wu , title =. Proc. Int. Conf. Learn. Represent. (. 2020 , publisher =

work page 2020
[27]

Martin Trimmel and Henning Petzka and Cristian Sminchisescu , title =. Proc. Int. Conf. Learn. Represent. (. 2021 , publisher =

work page 2021
[28]

Journal of Computational Mathematics , volume =

Juncai He and Lin Li and Jinchao Xu and Chunyue Zheng , title =. Journal of Computational Mathematics , volume =. 2020 , doi =

work page 2020
[29]

2023 , doi =

Christoph Hertrich and Amitabh Basu and Marco Di Summa and Martin Skutella , title =. 2023 , doi =

work page 2023
[30]

Rao , title =

Kuan-Lin Chen and Harinath Garudadri and Bhaskar D. Rao , title =. Advances in Neural Information Processing Systems , volume =

work page
[31]

Christian Haase and Christoph Hertrich and Georg Loho , title =. Proc. Int. Conf. Learn. Represent. (. 2023 , publisher =

work page 2023
[32]

Baraniuk , title =

Randall Balestriero and Romain Cosentino and Behnaam Aazhang and Richard G. Baraniuk , title =. Advances in Neural Information Processing Systems , volume =

work page
[33]

Bartlett , title =

Martin Anthony and Peter L. Bartlett , title =. 2002 , isbn =

work page 2002
[34]

Nature , volume =

Yann LeCun and Yoshua Bengio and Geoffrey Hinton , title =. Nature , volume =. 2015 , doi =

work page 2015
[35]

Hinton , title =

Vinod Nair and Geoffrey E. Hinton , title =. Proc. Int. Conf. Mach. Learn. (. 2010 , publisher =

work page 2010
[36]

Gennadiy Averkov and Christopher Hojny and Maximilian Merkert , title =. Proc. Int. Conf. Learn. Represent. (. 2025 , publisher =

work page 2025
[37]

Scikit-learn: Machine Learning in

Fabian Pedregosa and Ga. Scikit-learn: Machine Learning in. J. Mach. Learn. Res. , volume =. 2011 , doi =

work page 2011
[38]

Alex Krizhevsky , title =

work page
[39]

Hinton , title =

Alex Krizhevsky and Ilya Sutskever and Geoffrey E. Hinton , title =. Advances in Neural Information Processing Systems , volume =

work page
[40]

Gradient-Based Learning Applied to Document Recognition , journal =

Yann LeCun and L. Gradient-Based Learning Applied to Document Recognition , journal =. 1998 , doi =

work page 1998
[41]

Qinghua Tao and Li Li and Xiaolin Huang and Xiangming Xi and Shuning Wang and Johan A. K. Suykens , title =. Nature Reviews Methods Primers , volume =. 2022 , doi =

work page 2022
[42]

Laine , title =

Max Milkert and David Hyde and Forrest J. Laine , title =. Proc. Int. Conf. Mach. Learn. (. 2025 , publisher =

work page 2025
[43]

On the Local Complexity of Linear Regions in Deep

Niket Patel and Guido Mont. On the Local Complexity of Linear Regions in Deep. Proc. Int. Conf. Mach. Learn. (. 2025 , publisher =

work page 2025
[44]

Duc Anh Nguyen and Ernesto Araya and Adalbert Fono and Gitta Kutyniok , title =. Proc. Int. Conf. Mach. Learn. (. 2025 , publisher =

work page 2025
[45]

2022 , doi =

Qiang Hu and Hao Zhang and Feifei Gao and Chengwen Xing and Jianping An , title =. 2022 , doi =

work page 2022
[46]

International Conference on Neural Information Processing , pages=

Comparative analysis of the linear regions in ReLU and LeakyReLU networks , author=. International Conference on Neural Information Processing , pages=. 2023 , organization=

work page 2023
[47]

International Conference on Artificial Neural Networks , pages=

Empirical Study on the Effect of Residual Networks on the Expressiveness of Linear Regions , author=. International Conference on Artificial Neural Networks , pages=. 2023 , organization=

work page 2023
[48]

arXiv preprint arXiv:2310.18725 , year=

The Evolution of the Interplay Between Input Distributions and Linear Regions in Networks , author=. arXiv preprint arXiv:2310.18725 , year=

work page arXiv
[49]

Bibliography file sample.bib was not found during this compilation

work page