Pith. sign in

REVIEW 1 major objections 49 references

The standard CNN pipeline is a cross-lattice operator whose composition is a morphological opening in neither lattice because ReLU admits only a global adjoint.

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 · grok-4.3

2026-06-30 13:39 UTC pith:LJLC77WZ

load-bearing objection The paper models the full CNN pipeline as a non-idempotent cross-lattice operator under MMBB theory, which directly accounts for depth adding power, and identifies three concrete idempotent layer designs. the 1 major comments →

arxiv 2605.24608 v1 pith:LJLC77WZ submitted 2026-05-23 cs.AI cs.CVcs.LG

Lattice theory and algebraic models for deep convolutional learning based on mathematical morphology

Gustavo (Jesus) Angulo This is my paper
classification cs.AI cs.CVcs.LG
keywords lattice theorymathematical morphologyconvolutional neural networksmorphological openingsReLU activationidempotenceadjunctionstranslation-invariant operators
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 applies the Matheron-Maragos-Banon-Barrera representation theory layer by layer to show that linear convolution acts as an erosion in the Fourier inf-semilattice, ReLU acts as a lattice-join closing, and max-pooling acts as a dilation in the pointwise max-plus lattice. Their composition therefore fails to be an idempotent opening in either lattice. The upper adjoint of ReLU is shown to be the global operator that returns the input on non-negative functions and negative infinity otherwise, so no local erosion can pair with it. These facts together establish the algebraic reason depth produces new representational capacity rather than repeated application of the same map. The work also characterises three layer designs that do produce genuine idempotent openings and unifies max-pooling, strided convolution, and Laplacian pyramids under adjoint pyramid theory.

Core claim

The central claim is that the standard CNN pipeline (linear convolution + ReLU + flat max-pooling) is a cross-lattice operator: convolution is an erosion in the Fourier inf-semilattice, ReLU is a lattice-join closing, and max-pooling is a dilation in the pointwise max-plus lattice, so the composition is a morphological opening in neither lattice. The upper adjoint of ReLU in the pointwise lattice is the global operator that leaves globally non-negative functions unchanged and maps everything else to negative infinity; consequently no local morphological erosion forms an adjunction pair with ReLU. This non-idempotence of the composed layer is presented as the precise algebraic reason depth in

What carries the argument

Layer-wise application of the Matheron-Maragos-Banon-Barrera universal representation theory for translation-invariant operators, distinguishing the Fourier inf-semilattice from the pointwise max-plus lattice and tracking adjunctions and idempotence.

Load-bearing premise

The Matheron-Maragos-Banon-Barrera universal representation theory for translation-invariant operators applies systematically and without additional restrictions to every layer of a standard deep network.

What would settle it

A concrete local operator that forms the upper adjoint of ReLU inside the pointwise max-plus lattice, or an explicit example of a standard convolution-ReLU-max-pooling composition that is idempotent, would falsify the central claims.

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

If this is right

  • Because standard layers are not idempotent, stacking them changes the overall map and depth therefore increases representational capacity.
  • The three identified designs (pure max-plus morphological layer, spectral Wiener layer, self-dual morphological layer) each produce idempotent openings with associated fixed-point and convergence guarantees.
  • Max-pooling, strided convolution, and the Laplacian pyramid are instances of the same Goutsias-Heijmans adjoint pyramid construction.
  • The Activation-Pooling Dilation factorisation supplies the mathematically correct adjoint for the activation-plus-pooling block.

Where Pith is reading between the lines

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

  • Networks built from the three idempotent layer types might reach equivalent performance with fewer layers because each layer is a projection onto its fixed-point set.
  • The global character of ReLU's adjoint suggests checking whether other common activations possess local adjoints inside one of the two lattices.
  • The same cross-lattice analysis can be applied directly to residual blocks and encoder-decoder skip connections to determine whether they preserve or break idempotence.
  • Hybrid architectures that deliberately alternate between the two lattices could be used to control the degree of idempotence at each scale.

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

1 major / 0 minor

Summary. The paper develops a rigorous algebraic framework grounded in lattice theory and the Matheron-Maragos-Banon-Barrera (MMBB) universal representation theory for translation-invariant operators. It applies this systematically to model standard CNN pipelines (linear convolution + ReLU + flat max-pooling), ResNets, and encoder-decoder networks. The principal claims are that this pipeline is a cross-lattice operator (convolution as Fourier erosion, ReLU as pointwise closing, pooling as dilation) whose composition is a morphological opening in neither lattice, that ReLU's upper adjoint is a global (non-local) operator, and that this non-idempotence explains the representational power added by depth. The paper identifies three idempotent opening designs (pure max-plus, spectral Wiener, self-dual morphological), establishes fixed-point and convergence theory, and unifies max-pooling/strided convolution/Laplacian pyramid under Goutsias-Heijmans adjoint pyramid theory with an Activation-Pooling Dilation factorisation.

Significance. If the derivations hold, the work supplies a precise algebraic account of why depth confers genuine power in standard CNNs (non-idempotence arising from cross-lattice composition and global adjunction) rather than merely empirical observation. It also supplies concrete, fully characterised idempotent layer designs and a unification of pooling mechanisms, which could guide architecture search and theoretical analysis in morphological and convolutional learning.

major comments (1)
  1. [Abstract] Abstract: the central claims that the CNN pipeline is a cross-lattice operator whose composition is a morphological opening in neither lattice, and that ReLU possesses a global upper adjoint precluding any local erosion adjunction, are asserted without derivations, explicit operator definitions, or verification that MMBB conditions (translation invariance, lattice completeness) hold for the ReLU and pooling layers in the pointwise and Fourier lattices. This directly affects the load-bearing step that non-idempotence follows from the adjoint derivation.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful review and for highlighting the need for clearer linkage between the abstract claims and the supporting material. We address the single major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claims that the CNN pipeline is a cross-lattice operator whose composition is a morphological opening in neither lattice, and that ReLU possesses a global upper adjoint precluding any local erosion adjunction, are asserted without derivations, explicit operator definitions, or verification that MMBB conditions (translation invariance, lattice completeness) hold for the ReLU and pooling layers in the pointwise and Fourier lattices. This directly affects the load-bearing step that non-idempotence follows from the adjoint derivation.

    Authors: We agree that the abstract, being concise, states the principal results without reproducing the derivations or operator definitions. These are supplied in the body: explicit definitions of the layers as operators on the pointwise max-plus and Fourier inf-semilattices appear in Section 2; verification of the MMBB hypotheses (translation invariance of each layer and completeness of both lattices) is given in Sections 3–4; the cross-lattice composition, the explicit form of ReLU’s global upper adjoint, and the resulting non-idempotence are derived in Section 5. To make the abstract self-contained for readers, we will add brief parenthetical references to these sections and a short clause indicating that the non-idempotence follows directly from the adjoint calculation. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation applies external MMBB theory

full rationale

The paper grounds its analysis in the external Matheron--Maragos--Banon--Barrera universal representation theory for translation-invariant operators, applying it to identify convolution as Fourier erosion, ReLU as pointwise closing, and max-pooling as dilation. The cross-lattice non-idempotence and global upper adjoint of ReLU follow from these lattice definitions and adjunction properties without any reduction to fitted parameters, self-definitional loops, or load-bearing self-citations. The central claim that depth yields genuine power because the composition is not an opening in either lattice is derived from the imported theory rather than from internal data fits or renamed inputs. No equations or steps exhibit the forbidden patterns; the framework remains self-contained against the cited external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claims rest on the MMBB representation theorem for translation-invariant operators and standard lattice axioms for adjunctions and idempotence. No free parameters or invented entities are introduced in the abstract.

axioms (2)
  • domain assumption Matheron-Maragos-Banon-Barrera universal representation theory for translation-invariant operators
    Central tool stated to be applied systematically to every layer.
  • standard math Existence of upper and lower adjoints in the pointwise and Fourier lattices
    Used to classify ReLU and convolution as erosions or closings.

pith-pipeline@v0.9.1-grok · 5842 in / 1412 out tokens · 35997 ms · 2026-06-30T13:39:45.483335+00:00 · methodology

0 comments
read the original abstract

We develop a rigorous algebraic framework for deep convolutional architectures, CNNs, ResNets, and encoder--decoder networks such as UNet, grounded in lattice theory and mathematical morphology. The central tool is the Matheron--Maragos--Banon--Barrera (MMBB) universal representation theory for translation-invariant operators, which we apply systematically to every layer of a standard deep network. The principal finding is that the standard CNN pipeline (linear convolution~$+$ ReLU~$+$ flat max-pooling) is a cross-lattice operator: the convolution is an erosion in the Fourier inf-semilattice while ReLU is a lattice-join closing and max-pooling is a dilation in the pointwise max-plus lattice, and their composition is a morphological opening in neither. A second finding is that the upper adjoint of ReLU in the pointwise lattice is a global (non-local) operator, the identity on globally non-negative functions and $-\infty$ otherwise, so no local morphological erosion can form an adjunction pair with ReLU. These two results together provide the precise algebraic reason why depth in standard CNNs introduces genuine representational power: the composed layer is not idempotent. Three layer designs that are genuine idempotent openings are identified and fully characterised: the pure max-plus morphological layer (pointwise lattice), the spectral Wiener layer (Fourier lattice), and the self-dual morphological layer. We establish a complete fixed-point and convergence theory. The framework also unifies max-pooling, strided convolution, and the Laplacian pyramid under the Goutsias--Heijmans adjoint pyramid theory, and gives the Activation--Pooling Dilation (APD) factorisation with its correct adjoint.

Figures

Figures reproduced from arXiv: 2605.24608 by Gustavo (Jesus) Angulo.

Figure 1
Figure 1. Figure 1: MMBB activation pipeline and APD factorisation. Node colours denote the lattice each operator lives in: blue = Fourier (L n , ≤F ), orange = pointwise (L , ≤). The red arrow marks the cross-lattice jump from Fourier erosion to pointwise dilation. (a) Convolution + ReLU as an MMBB erosion pipeline (Corollary 6.11): Σ Spec K (Fourier) feeds basis erosions from B ±; their suprema combine, and the APD applies … view at source ↗
Figure 2
Figure 2. Figure 2: Positive/negative decomposition and MMBB activation shape. f +(x) = max(0, f) (orange) and f −(x) = max(0, −f) (teal) are both dila￾tions in (L , ≤); their difference recovers f = f + − f − (dashed diagonal, Re￾mark 6.10). The blue dotted line shows a MMBB morphological activation σ M B,c (Definition 6.8): an erosion-based operator with reduced slope and upper cap at level c, generalising ReLU [PITH_FULL_… view at source ↗
Figure 3
Figure 3. Figure 3: Three algebraic views of a CNN layer. Node colours denote the lat￾tice: blue = Fourier, orange = pointwise. Red arrows mark cross-lattice jumps; green arrows mark same-lattice adjoint pairs. (a) Standard CNN: the convolu￾tion ε Conv k lives in the Fourier lattice while δ ReLU and δ MP R live in the pointwise lattice; the red arrow marks the cross-lattice jump that breaks idempotency. (b) APD factorisation:… view at source ↗
Figure 4
Figure 4. Figure 4: Morphological interpretation of the ResNet block. (a) Standard block: residual F acts on f; skip connection adds f at ⊕. (b) Morphological reading (Proposition 7.6): when F ≈ γ M b − id, the block computes γ M b (f) = δb ∗ (εb(f)). The red-dashed skip carries the top-hat residue Γ(f) = f −γ M b (f), re￾injected before the adjoint dilation. The green loop expresses exact idempotency of the Type-I opening [… view at source ↗
Figure 5
Figure 5. Figure 5: Standard UNet (a) vs. morphological UResNet (b). Colours: or￾ange = pointwise lattice. (a) Encoder: strided convolutions with max-pooling (cross-lattice); decoder: transposed convolutions; skip connections (dashed) carry concatenated features. (b) Encoder: erosion-decimation ε ↓R bℓ ; decoder: adjoint dilation-interpolation δ ∗↑R bℓ (green arrows, forming the adjoint pair of Proposition 5.2); skip connecti… view at source ↗
Figure 6
Figure 6. Figure 6: Three CNN-like compositions and their lattice structure. Colours: blue = Fourier, orange = pointwise. Red arrows mark cross-lattice jumps; green arrows and loops mark same-lattice adjoint pairs and idempotency. (a) Stan￾dard CNN (Theorem 8.5): cross-lattice, not idempotent. (b) Type-I opening (Theorem 8.4(i)): both operators in (L , ≤); exactly idempotent. (c) Type-II opening (Theorem 8.4(ii)): both operat… view at source ↗
Figure 7
Figure 7. Figure 7: Convergence behaviour of Type-I layer iterations (Theorem 8.10). Orange nodes live in the pointwise lattice. (a) Iterating γ M b : the sequence is non-increasing and stabilises after one step (green arrow = idempotent). (b) Correct ResNet: when F ≈ γ −id, the block computes γ M b (f); stacking n blocks still converges in one step. (c) Naive residual Φ res(f) = γ(f) + f: sequence grows at least linearly in … view at source ↗
Figure 8
Figure 8. Figure 8: Structure of the median inf-semilattice and self-dual operators. Teal nodes live in (Fun(E, R), ⪯). (a) Median ordering: s ⪯ t iff same sign and |s| ≤ |t|; ⊥ = 0; opposite-sign elements are incomparable. (b) Binary infimum ⊓: value closest to zero (same sign). (c) Self-dual erosion ε Med W and adjoint dilation δ ∗Med W compose into γ Med W (green loop = exact idempotency; dashed arc = self-duality γ Med W … view at source ↗
Figure 9
Figure 9. Figure 9: Self-dual activation functions and symmetric pooling. (a) Standard ReLU (β − = 0): a closing in (L , ≤), not a dilation in (L , ⪯) since it zeros all negative values. Leaky ReLU (0 < β− < 1): a proper dilation in (L , ⪯), preserving sign and scaling negative values (Proposition 9.11(i)). Identity (β + = β − = 1, dashed teal): trivial self-dual dilation. (b) Symmetric max-pooling vs. self-dual opening γ Med… view at source ↗
Figure 10
Figure 10. Figure 10: Six activation functions and their morphological classification. Node colours denote the lattice: orange = pointwise, teal = median, blue = MMBB. ReLU (orange): closing in (L , ≤); global non-pointwise adjoint (Proposition 6.1). Leaky/PReLU (teal solid, β − = 0.15): dilation in (L , ⪯) (Proposition 9.11(i)). ELU and Softplus (gray): smooth approximations; no lattice characterisation. Absolute value |h| (t… 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

49 extracted references · 49 canonical work pages · 1 internal anchor

  1. [1]

    Some open questions on morphological operators and representations in the deep learning era,

    J. Angulo, “Some open questions on morphological operators and representations in the deep learning era,” inDiscrete Geometry and Mathematical Morphology (DGMM 2021), Lecture Notes in Computer Science, vol. 12708, pp. 3–19, Springer, 2021.(Invited keynote.)

    work page 2021

  2. [2]

    Nonlinear Representation Theory of Equivariant CNNs on Homogeneous Spaces Using Group Morphology,

    J. Angulo, “Nonlinear Representation Theory of Equivariant CNNs on Homogeneous Spaces Using Group Morphology,” inDiscrete Geometry and Mathematical Morphology (DGMM 2024), Lecture Notes in Computer Science, vol. 14605, pp. 255–267, Springer, 2024

    work page 2024

  3. [3]

    Group morphology fixed points on homogeneous spaces for deep learning equivariant networks,

    G. J. Angulo, “Group morphology fixed points on homogeneous spaces for deep learning equivariant networks,” inGeometric Science of Information (GSI 2025), Lecture Notes in Computer Science, vol. 15537, pp. 41–50, Springer, 2025

    work page 2025

  4. [4]

    A mathematical morphology view of the universal representation of scatter- ingnetworks,

    G. J. Angulo, “A mathematical morphology view of the universal representation of scatter- ingnetworks,” inDiscrete Geometry and Mathematical Morphology (DGMM 2025), Lecture Notes in Computer Science, vol. 16296, pp. 357–370, Springer, 2025

    work page 2025

  5. [5]

    Understanding deep neural net- works with rectified linear units,

    R. Arora, A. Basu, P. Mianjy, and A. Mukherjee, “Understanding deep neural net- works with rectified linear units,” inInternational Conference on Learning Representations (ICLR), 2018

    work page 2018

  6. [6]

    Decomposition of mappings between complete lattices by mathematical morphology, Part I: General lattices,

    G. J. F. Banon and J. Barrera, “Decomposition of mappings between complete lattices by mathematical morphology, Part I: General lattices,”Signal Processing, vol. 30, no. 3, pp. 299–327, 1993

    work page 1993

  7. [7]

    Morphological adjunctions rep- resented by matrices in max-plus algebra for signal and image processing,

    S. Blusseau, S. Velasco-Forero, J. Angulo, and I. Bloch, “Morphological adjunctions rep- resented by matrices in max-plus algebra for signal and image processing,” inDiscrete Geometry and Mathematical Morphology (DGMM 2022), Lecture Notes in Computer Sci- ence, vol. 13493, pp. 206–218, Springer, 2022

    work page 2022

  8. [8]

    Training morphological neural networks with gradient descent: some theoret- ical insights,

    S. Blusseau, “Training morphological neural networks with gradient descent: some theoret- ical insights,” inDiscrete Geometry and Mathematical Morphology (DGMM 2024), Lecture Notes in Computer Science, vol. 14605, pp. 228–240, Springer, 2024

    work page 2024

  9. [9]

    Improving mor- phological networks for learning image-to-image transforms,

    A. Bottenmuller, G. Tochon, R. Hermary, É. Puybareau, and J. Angulo, “Improving mor- phological networks for learning image-to-image transforms,” inDiscrete Geometry and Mathematical Morphology (DGMM 2025), Lecture Notes in Computer Science, vol. 16296, pp. 449–461, Springer, 2025

    work page 2025

  10. [10]

    Categorical foundations of gradient-based learning,

    G. S. H. Cruttwell, B. Gavranović, N. Ghani, P. Wilson, and F. Zanasi, “Categorical foundations of gradient-based learning,” inProgramming Languages and Systems – ESOP 2022, Lecture Notes in Computer Science, vol. 13240, pp. 1–28, Springer, 2022

    work page 2022

  11. [11]

    Theory of morphological neural networks,

    J. L. Davidson and G. X. Ritter, “Theory of morphological neural networks,” inProc. SPIE Digital Optical Computing II, vol. 1704, pp. 378–389, 1992

    work page 1992

  12. [12]

    Advances in morphological neural networks: training, pruning and enforcing shape constraints,

    N. Dimitriadis and P. Maragos, “Advances in morphological neural networks: training, pruning and enforcing shape constraints,” inIEEE ICASSP, pp. 3825–3829, 2021

    work page 2021

  13. [13]

    Morphological neural networks: expressing and learning better geometric features,

    N. Dimitriadis and P. Maragos, “Morphological neural networks: expressing and learning better geometric features,” inIEEE ICCV, 2023

    work page 2023

  14. [14]

    Learning morphological representations of image transformations: influence of initialization and layer differentiability,

    M. Dimitrova, S. Blusseau, and S. Velasco-Forero, “Learning morphological representations of image transformations: influence of initialization and layer differentiability,” inDiscrete Geometry and Mathematical Morphology (DGMM 2025), Lecture Notes in Computer Sci- ence, Springer, 2025

    work page 2025

  15. [15]

    Backprop as functor: a compositional perspec- tive on supervised learning,

    B. Fong, D. I. Spivak, and R. Tuyéras, “Backprop as functor: a compositional perspec- tive on supervised learning,” inProc. 34th Annual ACM/IEEE Symposium on Logic in Computer Science (LICS), pp. 1–13, 2019. LATTICE THEORY AND ALGEBRAIC MODELS FOR DEEP CONVOLUTIONAL NETWORKS 65

    work page 2019

  16. [16]

    Training deep morphological neural networks as universal approximators,

    K. Fotopoulos and P. Maragos, “Training deep morphological neural networks as universal approximators,” arXiv:2505.09710, 2025

    work page arXiv 2025

  17. [17]

    Deep morphological networks,

    G. Franchi, M. Fehri, and A. Yao, “Deep morphological networks,”Pattern Recognition, vol. 102, p. 107246, 2020

    work page 2020

  18. [18]

    Goodfellow, Y

    I. Goodfellow, Y. Bengio, and A. Courville,Deep Learning, MIT Press, 2016

    work page 2016

  19. [19]

    Nonlinear multiresolution signal decomposition schemes – Part I: morphological pyramids,

    J. Goutsias and H. J. A. M. Heijmans, “Nonlinear multiresolution signal decomposition schemes – Part I: morphological pyramids,”IEEE Transactions on Image Processing, vol. 9, no. 11, pp. 1862–1876, 2000

    work page 2000

  20. [20]

    The algebraic basis of mathematical morphology I. Dilations and erosions,

    H. J. A. M. Heijmans, C. Ronse. “The algebraic basis of mathematical morphology I. Dilations and erosions,”Computer Vision, Graphics, and Image Processing,vol. 50, no. 3, pp.245–295, 1990

    work page 1990

  21. [21]

    H. J. A. M. Heijmans,Morphological Image Operators, Academic Press, Boston, 1994

    work page 1994

  22. [22]

    Delving deep into rectifiers: surpassing human-level performance on ImageNet classification,

    K. He, X. Zhang, S. Ren, and J. Sun, “Delving deep into rectifiers: surpassing human-level performance on ImageNet classification,” inIEEE ICCV, pp. 1026–1034, 2015

    work page 2015

  23. [23]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in IEEE CVPR, pp. 770–778, 2016

    work page 2016

  24. [24]

    Learning grayscale mathematical morphology with smooth morphological layers,

    R. Hermary, G. Tochon, É. Puybareau, A. Kirszenberg, and J. Angulo, “Learning grayscale mathematical morphology with smooth morphological layers,”Journal of Mathematical Imaging and Vision, vol. 64, pp. 736–753, 2022

    work page 2022

  25. [25]

    A Morphological View on Traditional Signal Processing,

    R. Keshet, “A Morphological View on Traditional Signal Processing,” inMathematical Morphology and its Applications to Image and Signal Processing. Computational Imaging and Vision,vol. 18. Springer, 2002

    work page 2002

  26. [26]

    Implementation of linear digital filters based on mor- phological representation theory,

    M. Khosravi and R. W. Schafer, “Implementation of linear digital filters based on mor- phological representation theory,”IEEE Transactions on Signal Processing, vol. 42, no. 9, pp. 2264–2275, 1994

    work page 1994

  27. [27]

    Gradient-based learning applied to document recognition,

    Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,”Proceedings of the IEEE, vol. 86, no. 11, pp. 2278–2324, 1998

    work page 1998

  28. [28]

    Rectifier nonlinearities improve neural net- work acoustic models,

    A. L. Maas, A. Y. Hannun, and A. Y. Ng, “Rectifier nonlinearities improve neural net- work acoustic models,” inProc. ICML Workshop on Deep Learning for Audio, Speech and Language Processing, 2013

    work page 2013

  29. [29]

    Group invariant scattering,

    S. Mallat, “Group invariant scattering,”Communications on Pure and Applied Mathemat- ics, vol. 65, no. 10, pp. 1331–1398, 2012

    work page 2012

  30. [30]

    Morphological filters – Part I: their set-theoretic analysis and relations to linear shift-invariant filters,

    P. Maragos and R. W. Schafer, “Morphological filters – Part I: their set-theoretic analysis and relations to linear shift-invariant filters,”IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 35, no. 8, pp. 1153–1169, 1987

    work page 1987

  31. [31]

    Arepresentationtheoryformorphologicalimageandsignalprocessing,

    P.Maragos, “Arepresentationtheoryformorphologicalimageandsignalprocessing,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 11, no. 6, pp. 586–599, 1989

    work page 1989

  32. [32]

    Tropical geometry, morphological analysis, and deep neural networks,

    P. Maragos, “Tropical geometry, morphological analysis, and deep neural networks,”IEEE Signal Processing Magazine, vol. 38, no. 1, pp. 50–65, 2021

    work page 2021

  33. [33]

    The lattice overparameterization paradigm for the machine learningoflatticeoperators,

    D. Marcondes and J. Barrera, “The lattice overparameterization paradigm for the machine learningoflatticeoperators,” inDiscrete Geometry and Mathematical Morphology (DGMM 2024), Lecture Notes in Computer Science, vol. 14605, pp. 204–216, Springer, 2024

    work page 2024

  34. [34]

    Matheron,Random Sets and Integral Geometry, Wiley, New York, 1975

    G. Matheron,Random Sets and Integral Geometry, Wiley, New York, 1975

    work page 1975

  35. [35]

    On the number of linear regions of deep neural networks,

    G. F. Montufar, R. Pascanu, K. Cho, and Y. Bengio, “On the number of linear regions of deep neural networks,” inAdvances in Neural Information Processing Systems, vol. 27, pp. 2924–2932, 2014. 66 GUSTAVO (JESÚS) ANGULO

    work page 2014

  36. [36]

    Max-min representation of piecewise linear functions,

    S. Ovchinnikov, “Max-min representation of piecewise linear functions,”Beiträge zur Alge- bra und Geometrie, vol. 43, no. 1, pp. 297–302, 2002

    work page 2002

  37. [37]

    Group equivariant networks using morphological operators,

    V. Penaud-Polge, S. Velasco-Forero, and J. Angulo, “Group equivariant networks using morphological operators,” inDiscrete Geometry and Mathematical Morphology (DGMM 2024), Lecture Notes in Computer Science, vol. 14605, pp. 165–177, Springer, 2024

    work page 2024

  38. [38]

    Group equivariant morphological networks,

    V. Penaud-Polge, S. Velasco-Forero, and J. Angulo, “Group equivariant morphological networks,”SIAM Journal on Imaging Sciences, vol. 18, no. 4, pp. 2236–2276, 2025. DOI: 10.1137/24M1685766

    work page doi:10.1137/24m1685766 2025

  39. [39]

    Neural networks with hybrid morphological/rank/linear nodes: a unifying framework with applications to handwritten character recognition,

    L. F. C. Pessoa and P. Maragos, “Neural networks with hybrid morphological/rank/linear nodes: a unifying framework with applications to handwritten character recognition,”Pat- tern Recognition, vol. 33, no. 6, pp. 945–960, 2000

    work page 2000

  40. [40]

    On the spectral bias of neural networks,

    N. Rahaman et al., “On the spectral bias of neural networks,” inProc. 36th International Conference on Machine Learning, PMLR, vol. 97, pp. 5301–5310, 2019

    work page 2019

  41. [41]

    An introduction to morphological neural networks,

    G. X. Ritter and P. Sussner, “An introduction to morphological neural networks,” inProc. 13th ICPR, vol. 4, pp. 709–717, 1996

    work page 1996

  42. [42]

    U-Net: convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: convolutional networks for biomedical image segmentation,” inMedical Image Computing and Computer-Assisted Intervention (MICCAI), Lecture Notes in Computer Science, vol. 9351, pp. 234–241, Springer, 2015

    work page 2015

  43. [43]

    Scale equivariant neural net- works with morphological scale-spaces,

    M. Sangalli, S. Blusseau, S. Velasco-Forero, and J. Angulo, “Scale equivariant neural net- works with morphological scale-spaces,” inDiscrete Geometry and Mathematical Morphol- ogy (DGMM 2021), Lecture Notes in Computer Science, vol. 12708, pp. 483–495, Springer, 2021

    work page 2021

  44. [44]

    Serra,Image Analysis and Mathematical Morphology, Academic Press, London, 1982

    J. Serra,Image Analysis and Mathematical Morphology, Academic Press, London, 1982

    work page 1982

  45. [45]

    Attention is all you need,

    A. Vaswani et al., “Attention is all you need,” inAdvances in Neural Information Processing Systems, vol. 30, 2017

    work page 2017

  46. [46]

    Learnable empirical mode decomposition based on mathematical morphology,

    S. Velasco-Forero, R. Pagès, and J. Angulo, “Learnable empirical mode decomposition based on mathematical morphology,”SIAM Journal on Imaging Sciences, vol. 15, no. 1, pp. 23–44, 2022. DOI: 10.1137/21M1417867

    work page doi:10.1137/21m1417867 2022

  47. [47]

    Fixed point layers for geodesic morphologi- cal operations,

    S. Velasco-Forero, A. Rhim, and J. Angulo, “Fixed point layers for geodesic morphologi- cal operations,” in33rd British Machine Vision Conference (BMVC 2022), London, UK, November 21–24, 2022, BMVA Press, paper no. 480

    work page 2022

  48. [48]

    MorphoActivation: generalizing ReLU activation func- tion by mathematical morphology,

    S. Velasco-Forero and J. Angulo, “MorphoActivation: generalizing ReLU activation func- tion by mathematical morphology,” inDiscrete Geometry and Mathematical Morphology (DGMM 2022), Lecture Notes in Computer Science, vol. 13493, pp. 449–461, Springer, 2022

    work page 2022

  49. [49]

    Tropical Geometry of Deep Neural Networks

    L. Zhang, G. Naitzat, and L.-H. Lim, “Tropical geometry of deep neural networks,” arXiv:1805.07091, 2018. LATTICE THEORY AND ALGEBRAIC MODELS FOR DEEP CONVOLUTIONAL NETWORKS 67 Table 11.Standard deep learning architectures, their morphological interpre- tations established in this paper, and the morphologically-motivated variants proposed here. Architectu...

    work page internal anchor Pith review Pith/arXiv arXiv 2018