betaFeedback

WorksComputer Science

Two ReLU⁡\operatorname{ReLU} Layers is All You Need

Abstract

We prove that every ReLU⁡\operatorname{ReLU} network can be exactly simulated by a ReLU⁡\operatorname{ReLU} network with two hidden layers. We present two proofs: one that is based on fillings in homological algebra and a direct combinatorial proof.

Introduction

The ubiquity and success of neural networks as a model of computation motivates studying the capabilities of their architectures. This study raises familiar complexity-theoretic questions, like what is the class of functions they can compute, at what depth, with how many gates, etc.?

This work investigates the role of depth in neural networks. On the one hand, deep learning is one of the major successes of modern computer science. On the other hand, classical approximation theorems dating back to the late 1980s show that, for many activation functions including ReLU, a single hidden layer suffices to approximate any continuous function arbitrarily well on a compact domain [10, 19, 24]. Although various explanations for the power of compositions have been proposed [5, 26], a quantitative understanding of the role of depth remains incomplete [20, 17].

The role of depth has been studied from several perspectives. One line of work investigates depth-width tradeoffs in expressivity: can additional layers reduce the resources needed to represent or approximate a function [2, 12, 32]? Another asks whether depth can provide advantages for training [29]. Perhaps the most fundamental question, however, is whether deeper networks can represent functions that shallower networks cannot. Until now, this question has remained open.

We focus on neural networks that use the ReLU(xx) == max⁡{0,x}\max\{0, x\} activation function, which are among the most widely used ones in practice [13, 20]. A ReLU network with kk hidden layers (and depth k+1k + 1) is traditionally defined as the alternating composition

T(k)∘ReLU⁡∘T(k−1)∘⋯∘ReLU⁡∘T(1)∘ReLU⁡∘T(0)T^{(k)} \circ\operatorname{ReLU} \circ T^{(k-1)} \circ\cdots\circ\operatorname{ReLU} \circ T^{(1)} \circ\operatorname{ReLU} \circ T^{(0)}

where T(i)T^{(i)} is an affine map between two Euclidean spaces and ReLU stands for entry-wise application.

ReLU networks compute continuous and piecewise linear (CPWL) functions. These are continuous functions that are affine on each cell of a finite polyhedral partition of the input Euclidean space. Conversely, every CPWL function can be computed by a ReLU network (see e.g. [2]). Their ability to capture nonlinear behavior while retaining a simple local affine structure makes CPWL functions useful beyond neural networks. For example, they are central to finite element methods [16], and their structure enables exact verification of stability conditions in robotic control [11].

For a ring R⊆RR \subseteq\mathbb{R}, denote by CPWLn(R)\mathrm{CPWL}_n(R) the space of CPWL functions whose affine pieces have coefficients in RR. Denote by ReLUn,k(R)\mathrm{ReLU}_{n,k}(R) the subset of CPWLn(R)\mathrm{CPWL}_n(R) representable by ReLU networks with kk hidden layers, and weights and biases in RR. The obvious inclusions are

ReLUn,0(R)⊆ReLUn,1(R)⊆ReLUn,2(R)⊆⋯⊆CPWLn(R).(1)\mathrm{ReLU}_{n,0}(R) \subseteq\mathrm{ReLU}_{n,1}(R) \subseteq\mathrm{ReLU}_{n,2}(R) \subseteq\cdots\subseteq\mathrm{CPWL}_n(R). \tag*{(1)}

Our main results work when RR is any field containing Q\mathbb{Q}. We often write CPWLn\mathrm{CPWL}_n and ReLUn,k\mathrm{ReLU}_{n,k} for CPWLn(Q)\mathrm{CPWL}_n(\mathbb{Q}) and ReLUn,k(Q)\mathrm{ReLU}_{n,k}(\mathbb{Q}) respectively.

The main goal of this work is understanding the inclusions in the chain (1). Wang and Sun [34] identified a key structural result; see also [17, 4] for alternative, more geometric proofs. They showed that

MAX⁡n(x1,…,xn):=max⁡{x1,…,xn}\operatorname{MAX}_n(x_1,\ldots,x_n) := \max\{x_1,\ldots,x_n\}

is complete for CPWLn−1(R)\mathrm{CPWL}_{n-1}(\mathbb{R}) with respect to depth. That is, every function f∈CPWLn(R)f \in\mathrm{CPWL}_n(\mathbb{R}) can be expressed as a finite sum

f(x)=∑iσiMAX⁡n+1(Ai(x))(2)f(x) = \sum_i \sigma_i \operatorname{MAX}_{n+1}(A_i(x)) \tag*{(2)}

where Ai:Rn→Rn+1A_i:\mathbb{R}^n \to\mathbb{R}^{n+1} is an affine map and σi∈{±1}\sigma_i \in\{\pm1\}. The geometric proof by [4] for the universality of MAX⁡n+1\operatorname{MAX}_{n+1} relies on the fact that MAX⁡n+1\operatorname{MAX}_{n+1} is the support function of a simplex, and every polytope can be triangulated into simplices.

The first inclusion in (1) is trivially strict because ReLUn,0(R)\mathrm{ReLU}_{n,0}(\mathbb{R}) is precisely the set of affine functions. Mukherjee and Basu [27] showed that the second inclusion is strict for n≥3n \geq3 because MAX⁡3∈ReLU3,2(R)\operatorname{MAX}_3 \in\mathrm{ReLU}_{3,2}(\mathbb{R}) but MAX⁡3∉ReLU3,1(R)\operatorname{MAX}_3 \notin\mathrm{ReLU}_{3,1}(\mathbb{R}). In the other direction, Arora, Basu, Mianjy, and Mukherjee [2] observed that (2) implies that the length of the chain is at most logarithmic:

CPWLn=ReLUn,kfor k≥⌈log⁡2(n+1)⌉.\mathrm{CPWL}_n = \mathrm{ReLU}_{n,k} \quad\text{for } k \geq\left\lceil\log_2(n+1) \right\rceil.

The reason is that a divide-and-conquer network of depth ⌈log⁡2(n+1)⌉\left\lceil\log_2(n+1) \right\rceil computes MAX⁡n+1\operatorname{MAX}_{n+1}. Subsequently, Hertrich, Basu, Di Summa, and Skutella [17] conjectured that the length of the chain is exactly ⌈log⁡2(n+1)⌉\left\lceil\log_2(n+1) \right\rceil. In other words, they conjectured that MAX⁡n\operatorname{MAX}_n can not be computed by less than ⌈log⁡2(n)⌉\left\lceil\log_2(n) \right\rceil hidden layers.

A series of works validated this conjecture in a variety of special cases. Haase, Hertrich, and Loho [15] proved the conjecture over the integers:

k<⌈log⁡2(n)⌉⇒MAX⁡n∉ReLUn,k(Z).k < \left\lceil\log_2(n) \right\rceil\quad\Rightarrow\quad\operatorname{MAX}_n \notin\mathrm{ReLU}_{n,k}(\mathbb{Z}).

Averkov, Hojny, and Merkert [3] extended the logarithmic lower bounds to rings of the form Z[1/N]\mathbb{Z}[1/N] for an integer N>0N > 0. They showed that if a prime pp does not divide NN then

k<⌈log⁡p(n)⌉⇒MAX⁡n∉ReLUn,k(Z[1/N]).k < \left\lceil\log_p(n) \right\rceil\quad\Rightarrow\quad\operatorname{MAX}_n \notin\mathrm{ReLU}_{n,k}(\mathbb{Z}[1/N]).

There are more lower bounds for special classes of networks. Valerdi [33] proved the conjecture for input convex neural networks (ICNN), which was defined by Amos, Xu and Kolter [1] and studied in subsequent works (e.g. [7, 8, 9, 6, 18]). Grillo, Hertrich, and Loho [14] proved an Ω(log⁡log⁡n)\Omega(\log\log n) lower bound for ReLU networks that “respect the singularities of MAX⁡n\operatorname{MAX}_n”.

The conjecture was, however, recently disproved by Bakaev, Brunck, Hertrich, Stade, and Yehudayoff [4]. They proved that

MAX⁡5∈ReLU5,2\operatorname{MAX}_5 \in\mathrm{ReLU}_{5,2}

and concluded the asymptotic improvement

MAXn∈ReLUn,kfor k=⌈log⁡3(n−2)⌉+1.\mathrm{MAX}_n \in\mathrm{ReLU}_{n,k} \qquad\text{for } k = \lceil\log_{3}(n-2) \rceil+ 1.

Shortly after, computer-assisted constructions [28] showed that

MAX12∈ReLU12,2.\mathrm{MAX}_{12} \in\mathrm{ReLU}_{12,2}.

Again, this finite case can be converted into an improvement of the asymptotic bound from [4], yielding

MAXn∈ReLUn,kfor k=⌈log⁡6(n/2)⌉+1.\mathrm{MAX}_n \in\mathrm{ReLU}_{n,k} \qquad\text{for } k = \lceil\log_{6}(n/2) \rceil+ 1.

Our main result is that the entire depth hierarchy collapses to the second level, which is optimal by [27]. See Figure 1 for an illustration.

Figure 1. The depth hierarchy of ReLU networks collapses to its first 3 levels.

Theorem 1. For all n≥1n \ge1 and any field R⊆RR \subseteq\mathbb{R} containing Q\mathbb{Q},

CPWLn(R)=ReLUn,2(R).\mathrm{CPWL}_n(R) = \mathrm{ReLU}_{n,2}(R).

To prove our main result, it suffices to prove the inclusion

MAXn∈ReLUn,2(Q)\mathrm{MAX}_n \in\mathrm{ReLU}_{n,2}(\mathbb{Q})

We provide two proofs of the inclusion. The first is abstract and uses the language of homological algebra. The second is concrete and is based on exact combinatorial cancellations.

We, in fact, show that there is a ReLU network with two hidden layers for MAXn\mathrm{MAX}_n over the rationals where the precision required is poly⁡(n)\operatorname{poly}(n) many bits. The lower bound from [3] says that such a network must divide by the product of all primes up to ≈n\approx\sqrt{n}, which requires precision of ≈n\approx\sqrt{n} many bits. The combinatorial proof requires division by n!n!, which is consistent with this barrier. As written, the homological algebra proof also uses division by n!n!, but this is not strictly necessary. What the homological proof actually uses is division by 2k−12^k-1 for k=1,…,nk=1,\ldots,n, which is consistent with the barrier by Fermat’s little theorem.

Building the Simplex

Because MAXn\mathrm{MAX}_n is positively homogeneous, we can assume that all biases in the network expressing it are zero, see [17]. Because MAX2(x1,x2)=x1+ReLU⁡(x2−x1)\mathrm{MAX}_2(x_1,x_2)=x_1+\operatorname{ReLU}(x_2-x_1), we can and will work with networks that use MAX2\mathrm{MAX}_2 gates instead of ReLU gates, also known as maxout networks.

There is a duality between computing functions with MAX2\mathrm{MAX}_2 networks and building formal Minkowski differences of polytopes using the two operations: Minkowski sums and convex hulls of a union (see e.g. [17, 4] and references therein). This duality says that computing MAXn+1\mathrm{MAX}_{n+1} is equivalent to building the nn-dimensional simplex Δn\Delta_n. (There are n+1n+1 vertices and the dimension is nn.)

Computations with one hidden layer exactly correspond to differences of zonotopes, i.e. Minkowski sums of segments. Every zonotope can be decomposed into parallelepipeds ([31 Theorem (54)]) so one hidden layer is actually captured by parallelepipeds ∑i[ai,bi]\sum_i [a_i,b_i]. A general two-layer computation corresponds to the convex hull of two zonotopes.

Both of the constructions we describe use a particular class of two-layer components we call paramids. A paramid is the convex hull of an apex γ∈Rn\gamma\in\mathbb{R}^n and a parallelepiped ∑i[ai,bi]\sum_i [a_i,b_i]. For an example see Figure 2. The support function of a paramid is thus of the form

Figure 2. A pyramid given as the pyramid over a two-dimensional parallelepiped BB with apex qq.

MAX2(⟨γ,x⟩,∑iMAX2(⟨ai,x⟩,⟨bi,x⟩))\mathrm{MAX}_2\left(\langle\gamma,x\rangle,\sum_i \mathrm{MAX}_2(\langle a_i,x\rangle,\langle b_i,x\rangle)\right)

Paramids are the simplest two-layer polytopes and are thus a natural starting point to understand two-layer computations. Our results imply that ReLUn,2\mathrm{ReLU}_{n,2} is spanned by support functions of paramids. Actually, there is a more direct way to see this (see § 5.1). Informally, every convex hull of two zonotopes can be expressed algebraically in terms of paramids.

Another natural idea is to work modulo lower-dimensional pieces. A CPWL function f:Rn→Rf:\mathbb{R}^n \to\mathbb{R} is said to be lower-dimensional if there is a non-zero direction vv such that f(x+tv)=f(x)f(x+tv)=f(x) for all x∈Rnx \in\mathbb{R}^n and t∈Rt \in\mathbb{R}. So if, for each nn, we can write MAXn+1\mathrm{MAX}_{n+1} as a

MAXn+1=∑icihPi+R\mathrm{MAX}_{n+1}=\sum_i c_i h_{P_i}+R

where each PiP_i is a paramid and the remainder RR is a sum of lower-dimensional functions, then we are done by induction. The reason is that every lower-dimensional function is reducible to MAXn\mathrm{MAX}_n, which we already know how to compute.

Subdivisions give us a powerful tool in these arguments. If we can subdivide Δ\Delta into pieces that can be computed in kk hidden layers, then Δ\Delta can be computed by kk hidden layers. So subdivisions are “free” in the sense that they do not cost us anything in terms of depth. This follows from the following well-known inclusion-exclusion-like formula also known as valuation (see e.g. Lemma 6.1.1 in [30] or 5.20 in [25]):

Lemma 2 (Valuation property). Suppose that a polytope XX can be written as a union of polytopes X=A1∪⋯∪AkX = A_1 \cup\cdots\cup A_k. Then,

X+∑∅≠S⊆{1,…,k}∣S∣ even⋂i∈SAi=∑S⊆{1,…,k}∣S∣ odd⋂i∈SAi,X + \sum_{\substack{\varnothing\ne S \subseteq\{1,\ldots,k\} \\ |S|\ \mathrm{even}}} \bigcap_{i\in S} A_i = \sum_{\substack{S\subseteq\{1,\ldots,k\} \\ |S|\ \mathrm{odd}}} \bigcap_{i\in S} A_i,

where an empty polytope is interpreted as zero.

Our new goal is thus to build the simplex using paramids modulo sums of lower dimensional pieces. We present two such constructions. One construction proceeds by induction and uses the language of chain complexes and homological algebra. The second construction is based on a direct inclusion-exclusion formula.

Overview of the Homological Algebra Implementation

References

  1. [1]Brandon Amos, Lei Xu, and J. Zico Kolter. “Input Convex Neural Networks”. In: Proceedings of the 34th International Conference on Machine Learning. Vol. 70. Proceedings of Machine Learning Research. PMLR, 2017, pp. 146–155. URL: https://proceedings.mlr.press/v70/amos17b.html.arxiv.org/abs/1609.07152
  2. [2]Raman Arora, Amitabh Basu, Poorya Mianjy, and Anirbit Mukherjee. “Understanding Deep Neural Networks with Rectified Linear Units”. In: International Conference on Learning Representations. 2018. arXiv: 1611.01491. URL: https://openreview.net/forum?id=B1J_rgWRW.arxiv.org/abs/1611.01491
  3. [3]Gennadiy Averkov, Christopher Hojny, and Maximilian Merkert. “On the Expressiveness of Rational ReLU Neural Networks with Bounded Depth”. In: International Conference on Learning Representations. 2025. URL: https://openreview.net/forum?id=uREg3QHjLL.arxiv.org/abs/2502.06283
  4. [4]Egor Bakaev, Florestan Brunck, Christoph Hertrich, Jack Stade, and Amir Yehudayoff. “Better Neural Network Expressivity: Subdividing the Simplex”. In: Proceedings of the 58th Annual ACM Symposium on Theory of Computing. Preprint first posted in 2025. Salt Lake City, UT, USA: ACM, 2026, pp. 500–507. DOI: 10.1145/3798129.3800768. arXiv: 2505.14338. URL: https://arxiv.org/abs/2505.14338.
  5. [5]Yoshua Bengio. “Learning Deep Architectures for AI”. In: Foundations and Trends in Machine Learning 2.1 (2009), pp. 1–127. DOI: 10.1561/2200000006.DOI
  6. [6]Charlotte Bunne, Stefan G. Stark, Gabriele Gut, Jacobo Sarabia del Castillo, Mitch Levesque, Kjong-Van Lehmann, Lucas Pelkmans, Andreas Krause, and Gunnar Rätsch. “Learning Single-Cell Perturbation Responses Using Neural Optimal Transport”. In: Nature Methods 20.11 (2023), pp. 1759–1768. DOI: 10.1038/s41592-023-01969-x. URL: https://www.nature.com/articles/s41592-023-01969-x.DOI
  7. [7]Felix Büning, Adrian Schalbetter, Ahmed Aboudonia, Mathias Hudoba de Badyn, Philipp Heer, and John Lygeros. “Input Convex Neural Networks for Building MPC”. In: Proceedings of the 3rd Conference on Learning for Dynamics and Control. Vol. 144. Proceedings of Machine Learning Research. PMLR, 2021, pp. 251–262. URL: https://proceedings.mlr.press/v144/bunning21a.html.
  8. [8]Yize Chen, Yuanyuan Shi, and Baosen Zhang. “Optimal Control Via Neural Networks: A Convex Approach”. In: International Conference on Learning Representations. Preprint first posted in 2018; conference publication in 2019. 2019. arXiv: 1805.11835. URL: https://openreview.net/forum?id=H1MW72AcK7.arxiv.org/abs/1805.11835
  9. [9]Yize Chen, Yuanyuan Shi, and Baosen Zhang. “Data-Driven Optimal Voltage Regulation Using Input Convex Neural Networks”. In: Electric Power Systems Research 189 (2020), p. 106741. DOI: 10.1016/j.epsr.2020.106741. URL: https://www.sciencedirect.com/science/article/pii/S0378779620305447.DOI
  10. [10]George Cybenko. “Approximation by Superpositions of a Sigmoidal Function”. In: Mathematics of Control, Signals and Systems 2.4 (1989), pp. 303–314. DOI: 10.1007/BF02551274. URL: https://link.springer.com/article/10.1007/BF02551274.DOI
  11. [11]Hongkai Dai, Benoit Landry, Lujie Yang, Marco Pavone, and Russ Tedrake. “Lyapunov-Stable Neural-Network Control”. In: Proceedings of Robotics: Science and Systems. 2021. DOI: 10.15607/RSS.2021.XVII.063. URL: https://roboticsproceedings.org/rss17/p063.html.DOI
  12. [12]Ronen Eldan and Ohad Shamir. “The Power of Depth for Feedforward Neural Networks”. In: 29th Annual Conference on Learning Theory. Vol. 49. Proceedings of Machine Learning Research. PMLR, 2016, pp. 907–940. URL: https://proceedings.mlr.press/v49/eldan16.html.
  13. [13]Xavier Glorot, Antoine Bordes, and Yoshua Bengio. “Deep Sparse Rectifier Neural Networks”. In: Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics. Vol. 15. Proceedings of Machine Learning Research. PMLR, 2011, pp. 315–323. URL: https://proceedings.mlr.press/v15/glorot11a.html.
  14. [14]Moritz Grillo, Christoph Hertrich, and Georg Loho. “Depth-Bounds for Neural Networks via the Braid Arrangement”. In: Advances in Neural Information Processing Systems. Vol. 38. Curran Associates, Inc., 2025. arXiv: 2502.09324. URL: https://openreview.net/forum?id=X09fhSZkBh.arxiv.org/abs/2502.09324
  15. [15]Christian Alexander Haase, Christoph Hertrich, and Georg Loho. “Lower Bounds on the Depth of Integral ReLU Neural Networks via Lattice Polytopes”. In: International Conference on Learning Representations. 2023. arXiv: 2302.12553. URL: https://openreview.net/forum?id=2mvAL0AWaxY.arxiv.org/abs/2302.12553
  16. [16]Juncai He, Lin Li, Jinchao Xu, and Chunyue Zheng. “ReLU deep neural networks and linear finite elements”. In: Journal of Computational Mathematics 38.3 (2020), pp. 502–527.DOI
  17. [17]Christoph Hertrich, Amitabh Basu, Marco Di Summa, and Martin Skutella. “Towards Lower Bounds on the Depth of ReLU Neural Networks”. In: SIAM Journal on Discrete Mathematics 37.2 (2023). DOI: 10.1137/22M1489332.DOI
  18. [18]Christoph Hertrich and Georg Loho. Neural Networks and (Virtual) Extended Formulations. First posted in 2024; revised in 2025. 2024. DOI: 10.48550/arXiv.2411.03006. arXiv: 2411.03006. URL: https://arxiv.org/abs/2411.03006.
  19. [19]Kurt Hornik, Maxwell Stinchcombe, and Halbert White. “Multilayer Feedforward Networks Are Universal Approximators”. In: Neural Networks 2.5 (1989), pp. 359–366. DOI: 10.1016/0893-6080(89)90020-8. URL: https://www.sciencedirect.com/science/article/pii/0893608089900208.DOI
  20. [20]Joey Huchette, Gonzalo Muñoz, Thiago Serra, and Calvin Tsay. “When Deep Learning Meets Polyhedral Theory: A Survey”. In: INFORMS Journal on Computing (2026). Published online 12 March 2026; preprint first posted in 2023. DOI: 10.1287/ijoc.2024.0902. arXiv: 2305.00241. URL: https://pubsonline.informs.org/doi/10.1287/ijoc.2024.0902.DOI
  21. [21]Børge Jessen and Anders Thorup. “THE ALGEBRA OF POLYTOPES IN AFFINE SPACES”. In: Mathematica Scandinavica 43.2 (1978), pp. 211–240. ISSN: 00255521, 19031807. URL: http://www.jstor.org/stable/24491087 (visited on 10/11/2026).DOI
  22. [22]Christoph Koutschan, Bernhard Moser, Anton Ponomarchuk, and Josef Schicho. “Representing piecewise linear functions by functions with small arity”. In: Applicable Algebra in Engineering, Communication and Computing 36.4 (July 1, 2025), pp. 595–610. ISSN: 1432-0622. DOI: 10.1007/s00200-023-00627-1. URL: https://doi.org/10.1007/s00200-023-00627-1.
  23. [23]Christoph Koutschan, Anton Ponomarchuk, and Josef Schicho. “Representing Piecewise-Linear Functions by Functions With Minimal Arity”. In: IEEE Transactions on Information Theory 72.9 (2026), pp. 6825–6831. DOI: 10.1109/TIT.2026.3705562.DOI
  24. [24]Moshe Leshno, Vladimir Ya. Lin, Allan Pinkus, and Shimon Schocken. “Multilayer Feedforward Networks with a Nonpolynomial Activation Function Can Approximate Any Function”. In: Neural Networks 6.6 (1993), pp. 861–867. DOI: 10.1016/S0893-6080(05)80131-5.DOI
  25. [25]Peter McMullen and Rolf Schneider. “Valuations on convex bodies”. In: Convexity and its applications. Springer, 1983, pp. 170–247.DOI
  26. [26]Hrushikesh Mhaskar, Qianli Liao, and Tomaso Poggio. “When and Why Are Deep Networks Better Than Shallow Ones?” In: Proceedings of the AAAI Conference on Artificial Intelligence 31.1 (2017). DOI: 10.1609/aaai.v31i1.10913. URL: https://ojs.aaai.org/index.php/AAAI/article/view/10913.DOI
  27. [27]Anirbit Mukherjee and Amitabh Basu. Lower Bounds over Boolean Inputs for Deep Neural Networks with ReLU Gates. Tech. rep. TR17-190. Rehovot, Israel: Electronic Colloquium on Computational Complexity, 2017. DOI: 10.48550/arXiv.1711.03073. arXiv: 1711.03073. URL: https://arxiv.org/abs/1711.03073.
  28. [28]Kilian Rueß, Gennadiy Averkov, Florestan Brunck, Moritz Grillo, Christoph Hertrich, Georg Loho, Jack Stade, Moritz Stargalla, Matthew Suin, and Martin Winter. “Shallow ReLU network representations via exact linear algebra”. In: arXiv preprint arXiv:2607.21651 (2026). Version 2, 1 September 2026. DOI: 10.48550/arXiv.2607.21651. arXiv: 2607.21651. URL: https://arxiv.org/abs/2607.21651v2.
  29. [29]Itay Safran and Jason D. Lee. “Optimization-Based Separations for Neural Networks”. In: Proceedings of Thirty Fifth Conference on Learning Theory. Vol. 178. Proceedings of Machine Learning Research. PMLR, 2022, pp. 3–64. URL: https://proceedings.mlr.press/v178/safran22a.html.arxiv.org/abs/2112.02393
  30. [30]Rolf Schneider. “Valuations on convex bodies”. In: Convex Bodies: The Brunn–Minkowski Theory. Encyclopedia of Mathematics and its Applications. Cambridge University Press, 2013, pp. 329–368.
  31. [31]G. C. Shephard. “Combinatorial Properties of Associated Zonotopes”. In: Canadian Journal of Mathematics 26.2 (1974), pp. 302–321. DOI: 10.4153/CJM-1974-032-5.DOI
  32. [32]Matus Telgarsky. “Benefits of Depth in Neural Networks”. In: 29th Annual Conference on Learning Theory. Vol. 49. Proceedings of Machine Learning Research. PMLR, 2016, pp. 1517–1539. URL: https://proceedings.mlr.press/v49/telgarsky16.html.arxiv.org/abs/1602.04485
  33. [33]Juan L. Valderi. On Minimal Depth in Neural Networks. arXiv preprint arXiv:2402.15315. First posted in 2024; version 6 revised 22 September 2026. 2026. DOI: 10.48550/arXiv.2402.15315. arXiv: 2402.15315. URL: https://arxiv.org/abs/2402.15315v6.
  34. [34]Shuning Wang and Xusheng Sun. “Generalization of Hinging Hyperplanes”. In: IEEE Transactions on Information Theory 51.12 (2005), pp. 4425–4431. DOI: 10.1109/TIT.2005.859246.DOI
  35. [35]Charles A. Weibel. An Introduction to Homological Algebra. Cambridge Studies in Advanced Mathematics. Cambridge University Press, 1994.DOI

Paper details

vlnc:W_rzudwjq5
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents