pith. sign in

arxiv: 2601.17749 · v2 · submitted 2026-01-25 · 📡 eess.SP · cs.ET· cs.LG· cs.NE

Over-The-Air Extreme Learning Machines with XL Reception via Nonlinear Cascaded Metasurfaces

Kyriakos Stylianopoulos , Mattia Fabiani , Giulia Torcolacci , Davide Dardari , George C. Alexandropoulos This is my paper

Pith reviewed 2026-05-16 11:27 UTC · model grok-4.3

classification 📡 eess.SP cs.ETcs.LGcs.NE
keywords Extreme Learning MachinesMetasurfacesOver-the-Air ComputingXL-MIMOPhysical Layer Machine LearningBinary ClassificationStacked Intelligent Metasurfaces
0
0 comments X

The pith

Cascaded metasurfaces with a fixed nonlinear layer implement an extreme learning machine over the air.

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

This paper establishes that an extremely large MIMO receiver built from cascaded metasurfaces can perform extreme learning machine inference directly in the physical layer. The architecture uses one fixed nonlinear metasurface layer facing the channel and subsequent tunable linear layers to realize the ELM weights over the air, with training done in closed form. A sympathetic reader would care because this shifts machine learning computation into the wireless hardware, potentially enabling faster and more efficient inference in future communication systems without heavy digital processing. Numerical results show it matches digital models in the large element limit.

Core claim

In the XL regime of metasurface elements, the XL-MIMO-ELM system using stacked intelligent metasurfaces with a fixed nonlinear front layer and tunable linear layers achieves performance comparable to digital and idealized machine learning models for binary classification tasks across diverse datasets and wireless scenarios.

What carries the argument

Stacked Intelligent Metasurfaces (SIM) consisting of a fixed nonlinear front layer followed by tunable linear layers that approximate the trained ELM weights.

If this is right

  • The system performs binary classification completely over-the-air using a single RF chain after the metasurface stack.
  • Training of the metasurface responses occurs in closed form without iterative optimization.
  • Accuracy remains comparable to digital ELMs across multiple datasets and wireless channel conditions in the XL limit.
  • The approach shows that embedding learning capabilities directly into wireless hardware is feasible.

Where Pith is reading between the lines

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

  • This physical implementation could reduce receiver power consumption by avoiding full digital signal processing chains.
  • The cascaded nonlinear-linear structure might support other inference tasks if the layer responses are redesigned accordingly.
  • Integration with goal-oriented communications could allow direct physical-layer decisions without transmitting raw data.

Load-bearing premise

That densely packed cascaded metasurfaces with one fixed nonlinear front layer and tunable linear layers can accurately approximate the trained ELM weights in the physical domain under realistic wireless propagation.

What would settle it

A real-world experiment in an XL MIMO setup where the metasurface receiver's classification accuracy falls substantially below that of a digital ELM trained on identical data.

Figures

Figures reproduced from arXiv: 2601.17749 by Davide Dardari, George C. Alexandropoulos, Giulia Torcolacci, Kyriakos Stylianopoulos, Mattia Fabiani.

Figure 1
Figure 1. Figure 1: The proposed XL MIMO system for implementing [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Classification accuracy of two NL-CMS-ELM variations versus the number of elements [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Classification accuracy of two NL-CMS-ELM vari [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
read the original abstract

The recently envisioned goal-oriented communications paradigm calls for the application of inference on wirelessly transferred data via Machine Learning (ML) tools. An emerging research direction deals with the realization of inference ML models directly in the physical layer of Multiple-Input Multiple-Output (MIMO) systems, which, however, entails certain significant challenges. In this paper, leveraging the technology of programmable MetaSurfaces (MSs), we present an eXtremely Large (XL) MIMO system that acts as an Extreme Learning Machine (ELM) performing binary classification tasks completely Over-The-Air (OTA), which can be trained in closed form. The proposed system comprises a receiver architecture consisting of densely parallel placed diffractive layers of XL MSs, also known as Stacked Intelligent Metasurfaces (SIM), followed by a single reception radio-frequency chain. The front layer facing the XL MIMO channel consists of identical unit cells of a fixed NonLinear (NL) response, whereas the remaining layers of elements of tunable linear responses are utilized to approximate OTA the trained ELM weights. Our numerical investigations showcase that, in the XL regime of MS elements, the proposed XL-MIMO-ELM system achieves performance comparable to that of digital and idealized ML models across diverse datasets and wireless scenarios, thereby demonstrating the feasibility of embedding OTA learning capabilities into future wireless systems.

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.

Referee Report

2 major / 2 minor

Summary. The paper proposes an XL-MIMO receiver using stacked intelligent metasurfaces (SIM) to realize an Extreme Learning Machine (ELM) for over-the-air binary classification. A fixed nonlinear front layer combined with tunable linear layers approximates the closed-form trained ELM weights physically, with numerical results claiming performance comparable to digital and idealized ML models across datasets and wireless scenarios in the XL regime.

Significance. If the physical approximation of ELM weights holds under realistic propagation, the approach could enable low-latency, energy-efficient inference directly in the wireless physical layer for goal-oriented communications. The closed-form training of ELM weights is a clear strength that avoids iterative optimization and supports the feasibility claim.

major comments (2)
  1. [Numerical Results] Numerical investigations (abstract and results): comparable performance is reported without error bars, exact simulation parameters, or any metric quantifying the approximation error between the trained digital ELM weight matrix and the realized physical metasurface responses after tuning.
  2. [System Model] System model for cascaded layers: the claim that densely packed layers with one fixed NL front and tunable linear responses can accurately synthesize the ELM hidden-layer mapping relies on an unvalidated assumption that inter-layer propagation and metasurface tuning incur negligible residual phase/amplitude or diffraction errors relative to the ideal matrix.
minor comments (2)
  1. [System Model] Notation for metasurface responses and propagation matrices could be clarified with an explicit end-to-end transfer function equation linking the physical layers to the ELM weight matrix.
  2. [Numerical Results] Figure captions for performance curves should include the precise number of MS elements, SNR range, and dataset sizes used in each scenario.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help improve the clarity and rigor of our work on over-the-air ELM realization via stacked intelligent metasurfaces. We address each major comment below with specific revisions where feasible.

read point-by-point responses

  1. Referee: [Numerical Results] Numerical investigations (abstract and results): comparable performance is reported without error bars, exact simulation parameters, or any metric quantifying the approximation error between the trained digital ELM weight matrix and the realized physical metasurface responses after tuning.

    Authors: We agree that the numerical results would benefit from added rigor and transparency. In the revised manuscript we will include error bars on all performance plots (computed over 100 independent channel realizations), provide a dedicated table listing all exact simulation parameters (including metasurface element counts per layer, inter-layer distances, carrier frequency, and tuning resolution), and introduce a quantitative approximation-error metric defined as the normalized Frobenius norm ||W_phys - W_ELM||_F / ||W_ELM||_F between the physically realized response and the trained digital ELM weight matrix. These additions will be placed in the results section and will directly quantify the fidelity of the OTA implementation. revision: yes

  2. Referee: [System Model] System model for cascaded layers: the claim that densely packed layers with one fixed NL front and tunable linear responses can accurately synthesize the ELM hidden-layer mapping relies on an unvalidated assumption that inter-layer propagation and metasurface tuning incur negligible residual phase/amplitude or diffraction errors relative to the ideal matrix.

    Authors: The system model adopts the standard far-field cascaded propagation model for SIMs, which is appropriate for the XL regime where the large aperture suppresses diffraction relative to the element spacing. Our numerical results already show that the OTA performance closely matches the ideal digital ELM across multiple datasets and scenarios, providing indirect validation of the approximation under the stated conditions. We acknowledge that a more explicit treatment of non-idealities would strengthen the paper. In revision we will add a new subsection discussing the impact of small residual phase/amplitude errors and include supplementary Monte-Carlo simulations with 1-5% perturbation levels to demonstrate robustness; however, a full experimental validation of every propagation effect lies beyond the scope of the current numerical study. revision: partial

Circularity Check

0 steps flagged

No circularity: digital ELM training followed by physical approximation of weights

full rationale

The derivation trains ELM output weights in closed form on digital data, then tunes metasurface responses to approximate those weights OTA. Performance parity is shown via numerical simulation of the physical forward model against digital baselines. No step reduces a claimed prediction to a fitted input by construction, no self-citation chain bears the central result, and the approximation is treated as an engineering task rather than a definitional identity. The paper remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The architecture rests on the domain assumption that metasurface unit cells can be fabricated and programmed to realize the required fixed nonlinear and tunable linear responses at scale; no free parameters are explicitly fitted in the abstract description, and no new physical entities are postulated.

axioms (1)
  • domain assumption Programmable metasurfaces can realize fixed nonlinear responses in the front layer and tunable linear responses in subsequent layers that sufficiently approximate digital ELM weights
    Invoked to justify the physical implementation of the ELM hidden layer

pith-pipeline@v0.9.0 · 5564 in / 1193 out tokens · 60675 ms · 2026-05-16T11:27:43.254134+00:00 · methodology

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

29 extracted references · 29 canonical work pages · 1 internal anchor

  1. [1]

    Toward goal- oriented semantic communications: New metrics, framework, and open challenges,

    A. Li, S. Wu, S. Meng, R. Lu, S. Sun, and Q. Zhang, “Toward goal- oriented semantic communications: New metrics, framework, and open challenges,”IEEE Wireless Commun., vol. 31, no. 5, pp. 238–245, 2024

    work page 2024

  2. [2]

    Stacked intelligent metasurfaces for task-oriented semantic communications,

    G. Huang, J. An, Z. Yang, L. Gan, M. Bennis, and M. Debbah, “Stacked intelligent metasurfaces for task-oriented semantic communications,” arXiv preprint arXiv:2407.15053, 2024

    work page arXiv 2024

  3. [3]

    Over- the-air edge inference via metasurfaces-integrated artificial neural net- works,

    K. Stylianopoulos, P. Di Lorenzo, and G. C. Alexandropoulos, “Over-the- air edge inference via metasurfaces-integrated artificial neural networks,” arXiv preprint arXiv:2504.00233, 2025

    work page arXiv 2025

  4. [4]

    A survey on over-the-air computation,

    A. ¸ Sahin and R. Yang, “A survey on over-the-air computation,”IEEE Commun. Surveys & Tuts., vol. 25, no. 3, pp. 1877–1908, 2023

    work page 1908

  5. [5]

    All-optical machine learning using diffractive deep neural networks,

    X. Lin, Y . Rivenson, N. T. Yardimci, M. Veli, Y . Luo, M. Jarrahi, and A. Ozcan, “All-optical machine learning using diffractive deep neural networks,”Science, vol. 361, no. 6406, pp. 1004–1008, 2018

    work page 2018

  6. [6]

    Electromagnetic wave-based extreme deep learning with nonlinear time-floquet entanglement,

    A. Momeni and R. Fleury, “Electromagnetic wave-based extreme deep learning with nonlinear time-floquet entanglement,”Nature Commun., vol. 13, no. 1, p. 2651, May 2022

    work page 2022

  7. [7]

    Stacked intelligent metasurface performs a 2D DFT in the wave domain for DOA estimation,

    J. An, C. Yuen, Y . L. Guan, M. Di Renzo, M. Debbah, and H. V . Poor, “Stacked intelligent metasurface performs a 2D DFT in the wave domain for DOA estimation,” inProc. IEEE Int. Commun. Conf., Denver, USA, 2024

    work page 2024

  8. [8]

    Implementing neural net- works over-the-air via reconfigurable intelligent surfaces,

    M. Hua, C. Bian, H. Wu, and D. Gündüz, “Implementing neural net- works over-the-air via reconfigurable intelligent surfaces,”arXiv preprint arXiv:2508.01840, 2025

    work page arXiv 2025

  9. [9]

    Over-the-air semantic alignment with stacked intelligent metasurfaces,

    M. E. Pandolfo, K. Stylianopoulos, G. C. Alexandropoulos, and P. Di Lorenzo, “Over-the-air semantic alignment with stacked intelligent metasurfaces,”arXiv preprint arXiv:2512.05657, 2025

    work page arXiv 2025

  10. [10]

    Metasurfaces-integrated wireless neural networks for lightweight over-the-air edge inference,

    K. Stylianopoulos, M. E. Pandolfo, P. Di Lorenzo, and G. C. Alexandropoulos, “Metasurfaces-integrated wireless neural networks for lightweight over-the-air edge inference,”IEEE Wireless Commun., under review, 2025

    work page 2025

  11. [11]

    Universal Approximation with XL MIMO Systems: OTA Classification via Trainable Analog Combining

    K. Stylianopoulos and G. C. Alexandropoulos, “Universal approxima- tion with XL MIMO systems: OTA classification via trainable analog combining,”arXiv preprint:2504.12758, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  12. [12]

    Extreme learning machine: Theory and applications,

    G.-B. Huang, Q.-Y . Zhu, and C.-K. Siew, “Extreme learning machine: Theory and applications,”Neurocomput., vol. 70, no. 1, pp. 489–501, 2006

    work page 2006

  13. [13]

    Fully complex extreme learning machine,

    M.-B. Li, G.-B. Huang, P. Saratchandran, and N. Sundararajan, “Fully complex extreme learning machine,”Neurocomput., vol. 68, pp. 306– 314, 2005

    work page 2005

  14. [14]

    Incremental extreme learning machine with fully complex hidden nodes,

    G.-B. Huang, M.-B. Li, L. Chen, and C.-K. Siew, “Incremental extreme learning machine with fully complex hidden nodes,”Neurocomput., vol. 71, no. 4, pp. 576–583, 2008

    work page 2008

  15. [15]

    Nonlinear EM-based signal processing,

    M. Fabiani, G. Torcolacci, and D. Dardari, “Nonlinear EM-based signal processing,” inProc. Asilomar Conf. Signals, Sys., and Comput., Pacific Grove, USA, 2025

    work page 2025

  16. [16]

    Nonlinear stacked intelligent surfaces for wireless systems,

    O. Abbas, A. Zayat, L. Markley, and A. Chaaban, “Nonlinear stacked intelligent surfaces for wireless systems,”arXiv preprint arXiv:2510.23780, Oct. 2025

    work page arXiv 2025

  17. [17]

    Approximation by superpositions of a sigmoidal function,

    G. Cybenko, “Approximation by superpositions of a sigmoidal function,” Math. Control Signal Sys., vol. 2, p. 303–314, 1989

    work page 1989

  18. [18]

    A theory of goal-oriented communication,

    O. Goldreich, B. Juba, and M. Sudan, “A theory of goal-oriented communication,”J. Assoc. Comput. Mach., vol. 59, no. 2, 2012

    work page 2012

  19. [19]

    Unitary evolution recurrent neural networks,

    M. Arjovsky, A. Shah, and Y . Bengio, “Unitary evolution recurrent neural networks,” inProc. Int. Conf. Mach. Learn., New York, USA, 2016

    work page 2016

  20. [20]

    Stacked intelligent metasurfaces for efficient holographic MIMO communications in 6G,

    J. An, C. Xu, D. W. K. Ng, G. C. Alexandropoulos, C. Huang, C. Yuen, and L. Hanzo, “Stacked intelligent metasurfaces for efficient holographic MIMO communications in 6G,”IEEE J. Sel. Areas Commun., vol. 41, no. 8, pp. 2380–2396, 2023

    work page 2023

  21. [21]

    PhysFad: Physics-based end-to-end channel modeling of RIS-parametrized environments with adjustable fading,

    R. Faqiri, C. Saigre-Tardif, G. C. Alexandropoulos, N. Shlezinger, M. F. Imani, and P. del Hougne, “PhysFad: Physics-based end-to-end channel modeling of RIS-parametrized environments with adjustable fading,” IEEE Trans. Wireless Commun., vol. 22, no. 1, pp. 580–595, 2023

    work page 2023

  22. [22]

    On the tacit linearity assumption in common cascaded models of RIS-parametrized wireless channels,

    A. Rabault, L. Le Magoarou, J. Sol, G. C. Alexandropoulos, N. Shlezinger, H. V . Poor, and P. del Hougne, “On the tacit linearity assumption in common cascaded models of RIS-parametrized wireless channels,”IEEE Trans. Wireless Commun., vol. 23, no. 8, pp. 10 001– 10 014, 2024

    work page 2024

  23. [23]

    Multilayer feedforward networks with a nonpolynomial activation function can approximate any function,

    M. Leshno, V . Y . Lin, A. Pinkus, and S. Schocken, “Multilayer feedforward networks with a nonpolynomial activation function can approximate any function,”Neural Netw., vol. 6, no. 6, pp. 861–867, 1993

    work page 1993

  24. [24]

    The universal approximation theorem for complex- valued neural networks,

    F. V oigtlaender, “The universal approximation theorem for complex- valued neural networks,”Appl. Comput. Harmon. Anal., vol. 64, pp. 33–61, 2023

    work page 2023

  25. [25]

    UCI machine learning repository,

    D. Dua and C. Graff, “UCI machine learning repository,” University of California, Irvine, School of Information and Computer Sciences,

    work page

  26. [26]

    Available: http://archive.ics.uci.edu/ml

    [Online]. Available: http://archive.ics.uci.edu/ml

    work page

  27. [27]

    The MNIST database of handwritten digit images for machine learning research,

    L. Deng, “The MNIST database of handwritten digit images for machine learning research,”IEEE Signal Process. Mag., vol. 29, no. 6, pp. 141– 142, 2012

    work page 2012

  28. [28]

    Dynamic metasurface antennas for 6G extreme massive mimo communications,

    N. Shlezinger, G. C. Alexandropoulos, M. F. Imani, Y . C. Eldar, and D. R. Smith, “Dynamic metasurface antennas for 6G extreme massive mimo communications,”IEEE Wireless Commun., vol. 28, no. 2, pp. 106–113, 2021

    work page 2021

  29. [29]

    Analytic framework for the effective rate of MISO fading channels,

    M. Matthaiou, G. C. Alexandropoulos, H. Q. Ngo, and E. G. Larsson, “Analytic framework for the effective rate of MISO fading channels,” IEEE Trans. Commun., vol. 60, no. 6, pp. 1741–1751, 2012

    work page 2012