pith. sign in

arxiv: 2605.26848 · v1 · pith:LH7NUSBGnew · submitted 2026-05-26 · ⚛️ physics.optics

Design principles for optoelectronic light-scattering reservoir computing at the edge of chaos

Geon Kim , YongKeun Park This is my paper

Pith reviewed 2026-06-29 16:13 UTC · model grok-4.3

classification ⚛️ physics.optics
keywords reservoir computinglight scatteringedge of chaosLyapunov exponentMackey-Glassspeech classificationoptical reservoirphysical computing
0
0 comments X

The pith

Three control axes in a light-scattering reservoir reach optima at the edge of chaos for sequential inference.

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

The authors systematically map reservoir dynamics, input-reservoir coupling, and reservoir interconnectivity to locate quantitative optima in an optoelectronic light-scattering system. Memory capacity peaks when the maximal Lyapunov exponent is near zero, and this location is reproduced in simulation. At the combined three-axis optimum the reservoir performs stable free-running Mackey-Glass prediction and reaches 84.5 percent accuracy on blind ten-class spoken-digit classification. A sympathetic reader would care because the rules link concrete hardware knobs to criticality concepts that could reduce training costs in physical sequential processors.

Core claim

By varying the three physical control axes the study identifies a quantitative optimum where memory capacity peaks at a near-zero maximal Lyapunov exponent. Operating at this point the reservoir supports stable free-running Mackey-Glass time-series prediction and achieves 84.5 percent blind classification accuracy on the 10-class Speech Commands benchmark. The design principles are stated in substrate-specific units yet rest on substrate-independent notions of criticality and balanced coupling.

What carries the argument

Three physical control axes—reservoir dynamics, input-reservoir coupling, and reservoir interconnectivity—whose joint optimum places the system at the edge of chaos and produces the reported performance.

If this is right

  • Stable Mackey-Glass prediction occurs in free-running mode at the three-axis optimum.
  • 84.5 percent blind accuracy is reached on the 10-class Speech Commands benchmark.
  • The memory-capacity peak is reproduced quantitatively by numerical simulation.
  • Input coupling shows a density-magnitude trade-off and interconnectivity shows an intermediate optimum.

Where Pith is reading between the lines

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

  • The same three-axis mapping could be applied to other reconfigurable photonic substrates to test whether the edge-of-chaos location remains optimal.
  • The density-magnitude trade-off in coupling strength may set practical limits on input-signal scaling in free-space optical reservoirs.
  • An intermediate interconnectivity optimum may turn out to be a general rule for balancing memory and nonlinearity across physical reservoir types.

Load-bearing premise

The memory-capacity peak at near-zero Lyapunov exponent is what causes the benchmark performance rather than merely correlating with it.

What would settle it

Fix the input-coupling and interconnectivity axes at their reported optima, deliberately shift the maximal Lyapunov exponent away from zero, and check whether prediction error rises or classification accuracy falls below 84.5 percent.

Figures

Figures reproduced from arXiv: 2605.26848 by Geon Kim, YongKeun Park.

Figure 1
Figure 1. Figure 1: Light-scattering reservoir computing platform and its three-axis design space. (a) Conceptual illustration of a recurrent neural network (RNN) or reservoir computing system. At each time 𝑡, the reservoir state vector 𝐡(𝑡) is obtained by the input vector 𝐱(𝑡) and previous state 𝐡(𝑡 − 1), which are rearranged to 𝐡(𝑡) through 𝐖in and 𝐖rec, respectively. (b) Simplified schematic of an optoelectronic reservoir … view at source ↗
Figure 2
Figure 2. Figure 2: Memory peaking at the dynamic criticality of the light scattering reservoir. [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Input–reservoir coupling and its memory-capacity optimum. (a) Encoding scheme for the input vector 𝐱(𝑡). Both the value and ratio of nonzero entries of 𝐖in are controlled. An input of dimension 𝐿 is mapped onto the display super-pixels that correspond to the nonzero 𝐖in entries repeatedly (blue arrows). (b) MC at the camera exposure closest to the dynamic criticality (MCcrit) as a function of the value (𝑤-… view at source ↗
Figure 4
Figure 4. Figure 4: Reservoir scale and scatterer interconnectivity exhibit a common saturating [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Experimental chaotic time-series prediction across three reservoir-dynamics regimes. The light scattering reservoir of size 10,000, configured at the input-reservoir coupling and interconnectivity identified in Figures 3 and 4, is operated at three detector exposure settings that place the dynamic (a) stable (𝜆max < 0), (b) critical (𝜆max ≈ 0), and (c) chaotic (𝜆max > 0) regimes. For each regime, the left … view at source ↗
Figure 6
Figure 6. Figure 6: Real-world spoken-digit recognition with the high-memory light scattering reservoir. (a) Input data processing pipeline for a spoken word recording data. Each recording is preprocessed into a 32-channel log-mel spectrogram, which is streamed along time. (b) Schematic description of spoken-digit recognition task; the goal is to accurately indicate which digit is uttered, among the 10 digits. Each recording … view at source ↗
read the original abstract

Physical reservoir computing offers an energy-efficient route to sequential cognitive inference by outsourcing nonlinear temporal mixing to hardware substrates with rich intrinsic dynamics, with free-space light-scattering systems particularly attractive for their parallelism and reconfigurability-yet practical design principles linking hardware control variables to computational performance have remained unestablished. Here, we establish such principles by systematically mapping three physical control axes of a reconfigurable optoelectronic light-scattering reservoir-reservoir dynamics, input-reservoir coupling, and reservoir interconnectivity-and identifying a quantitative optimum along each axis. Within this design landscape, we observe a memory-capacity peak that coincides with near-zero maximal Lyapunov exponent and is quantitatively reproduced in numerical simulation, extending edge-of-chaos confirmations previously reported in ion-gating and spin-wave reservoirs into the photonic substrate. The two remaining axes exhibit a density-magnitude trade-off in input coupling and an intermediate optimum in reservoir interconnectivity. Operating at the resulting three-axis optimum, the reservoir achieves stable Mackey-Glass chaotic time-series prediction in free-running mode and 84.5% blind classification accuracy on the 10-class Speech Commands spoken-digit benchmark; the principles, stated in substrate-specific units yet rooted in substrate-independent concepts of criticality and balanced coupling, provide a transferable framework for reconfigurable optical reservoir hardware.

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

1 major / 0 minor

Summary. The paper claims to establish practical design principles for a reconfigurable optoelectronic light-scattering reservoir by systematically mapping three physical control axes (reservoir dynamics via maximal Lyapunov exponent, input-reservoir coupling, and reservoir interconnectivity). It reports a memory-capacity peak coinciding with near-zero maximal Lyapunov exponent that is quantitatively reproduced in simulation, plus stable free-running Mackey-Glass prediction and 84.5% blind accuracy on the 10-class Speech Commands benchmark when operating at the identified three-axis optimum. The principles are presented as substrate-specific yet rooted in substrate-independent notions of criticality and balanced coupling.

Significance. If the reported optima and performance metrics hold under the stated conditions, the work supplies concrete, transferable design rules for photonic reservoir hardware that extend prior edge-of-chaos observations from other physical substrates. The quantitative experimental-simulation match on the memory-capacity peak and the free-running chaotic prediction constitute clear strengths.

major comments (1)
  1. [Abstract] Abstract: the central claim that performance occurs at the 'resulting three-axis optimum' and that the work 'extends edge-of-chaos confirmations' rests on the untested assumption that the near-zero Lyapunov-exponent condition contributes causally rather than coinciding incidentally with the optima identified on the input-coupling and interconnectivity axes. No direct comparison is described that isolates the contribution of the dynamics axis (e.g., performance when the other two axes are optimized while the Lyapunov exponent is held away from zero).

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the detailed and constructive report. The single major comment is addressed point-by-point below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that performance occurs at the 'resulting three-axis optimum' and that the work 'extends edge-of-chaos confirmations' rests on the untested assumption that the near-zero Lyapunov-exponent condition contributes causally rather than coinciding incidentally with the optima identified on the input-coupling and interconnectivity axes. No direct comparison is described that isolates the contribution of the dynamics axis (e.g., performance when the other two axes are optimized while the Lyapunov exponent is held away from zero).

    Authors: We agree that the manuscript does not contain a controlled isolation experiment in which input-coupling and interconnectivity are held at their identified optima while the maximal Lyapunov exponent is deliberately moved away from zero. The evidence presented consists of (i) an independent sweep of the dynamics axis that locates a memory-capacity peak at near-zero MLE, quantitatively reproduced in simulation, and (ii) the joint performance obtained when all three axes are set to their individually mapped optima. This constitutes a correlation rather than a direct causal isolation. We will therefore revise the abstract (and the corresponding discussion paragraph) to replace language implying causality with language that reports the observed coincidence and the resulting performance metrics. revision: partial

Circularity Check

0 steps flagged

No circularity: derivation chain maps control axes to observed optima without reduction to inputs by construction

full rationale

The paper systematically varies three hardware axes (reservoir dynamics, input coupling, interconnectivity), reports a memory-capacity peak coinciding with near-zero Lyapunov exponent (reproduced in simulation), and states performance at the resulting optimum. No equations, fitted-parameter renamings, or self-citation chains are exhibited that define the claimed optima or performance metrics in terms of themselves. The edge-of-chaos extension cites prior work on other substrates without the present authors' prior results serving as the sole justification for the photonic case. The central claims remain independent of the reported metrics and are not forced by definition or statistical fitting of the target quantities.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only; no free parameters, axioms, or invented entities can be extracted beyond the general assumption that Lyapunov exponent controls memory capacity.

pith-pipeline@v0.9.1-grok · 5754 in / 1167 out tokens · 27771 ms · 2026-06-29T16:13:24.098289+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

41 extracted references · 2 canonical work pages · 2 internal anchors

  1. [1]

    Brown, B

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, D. Amodei, in ...

    2020

  2. [2]

    Strubell, A

    E. Strubell, A. Ganesh, A. McCallum, in Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Association For Computational Linguistics, Florence, Italy, 2019, pp. 3645–3650

    2019

  3. [3]

    Carbon Emissions and Large Neural Network Training

    D. Patterson, J. Gonzalez, Q. Le, C. Liang, L.-M. Munguia, D. Rothchild, D. So, M. Texier, J. Dean, 2021, DOI 10.48550/arXiv.2104.10350

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2104.10350 2021

  4. [4]

    M. C. Mackey, L. Glass, Science 1977, 197, 287

    1977

  5. [5]

    G. Chen, C. Parada, G. Heigold, in 2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, Florence, Italy, 2014, pp. 4087–4091

    2014

  6. [6]

    Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition

    P. Warden, 2018, DOI 10.48550/ARXIV.1804.03209

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1804.03209 2018

  7. [7]

    M. Yan, C. Huang, P. Bienstman, P. Tino, W. Lin, J. Sun, Nat Commun 2024, 15, 2056

    2024

  8. [8]

    Wetzstein, A

    G. Wetzstein, A. Ozcan, S. Gigan, S. Fan, D. Englund, M. Soljačić, C. Denz, D. A. B. Miller, D. Psaltis, Nature 2020, 588, 39

    2020

  9. [9]

    Nishioka, T

    D. Nishioka, T. Tsuchiya, W. Namiki, M. Takayanagi, M. Imura, Y. Koide, T. Higuchi, K. Terabe, Sci. Adv. 2022, 8, eade1156

    2022

  10. [10]

    Nishimura, D

    M. Nishimura, D. Nishioka, W. Namiki, S. Hikasa, R. Iguchi, K. Terabe, T. Tsuchiya, Jpn. J. Appl. Phys. 2026, 65, 080903

    2026

  11. [11]

    Loeffler, R

    A. Loeffler, R. Zhu, J. Hochstetter, M. Li, K. Fu, A. Diaz-Alvarez, T. Nakayama, J. M. Shine, Z. Kuncic, Front. Neurosci. 2020, 14, 184

    2020

  12. [12]

    Larger, M

    L. Larger, M. C. Soriano, D. Brunner, L. Appeltant, J. M. Gutierrez, L. Pesquera, C. R. Mirasso, I. Fischer, Opt. Express 2012, 20, 3241

    2012

  13. [13]

    Brunner, M

    D. Brunner, M. C. Soriano, C. R. Mirasso, I. Fischer, Nat Commun 2013, 4, 1364

    2013

  14. [14]

    Vandoorne, P

    K. Vandoorne, P. Mechet, T. Van Vaerenbergh, M. Fiers, G. Morthier, D. Verstraeten, B. Schrauwen, J. Dambre, P. Bienstman, Nat Commun 2014, 5, 3541

    2014

  15. [15]

    D. Wang, Y. Nie, G. Hu, H. K. Tsang, C. Huang, Nat Commun 2024, 15, 10841

    2024

  16. [16]

    J. Dong, S. Gigan, F. Krzakala, G. Wainrib, in 2018 IEEE Statistical Signal Processing Workshop (SSP), IEEE, Freiburg, 2018, pp. 448–452

    2018

  17. [17]

    J. Dong, M. Rafayelyan, F. Krzakala, S. Gigan, IEEE J. Select. Topics Quantum Electron. 2020, 26, 1

    2020

  18. [18]

    Rafayelyan, J

    M. Rafayelyan, J. Dong, Y. Tan, F. Krzakala, S. Gigan, Phys. Rev. X 2020, 10, 041037

    2020

  19. [19]

    Jaeger, Bonn, Germany: German national research center for information technology gmd technical report 2001, 148

    H. Jaeger, Bonn, Germany: German national research center for information technology gmd technical report 2001, 148

    2001

  20. [20]

    Maass, T

    W. Maass, T. Natschläger, H. Markram, Neural Computation 2002, 14, 2531

    2002

  21. [21]

    Lukoševičius, in Neural Networks: Tricks of the Trade (Eds.: G

    M. Lukoševičius, in Neural Networks: Tricks of the Trade (Eds.: G. Montavon, G. B. Orr, K.-R. Müller), Springer Berlin Heidelberg, Berlin, Heidelberg, 2012, pp. 659–686

    2012

  22. [22]

    C. G. Langton, Physica D: Nonlinear Phenomena 1990, 42, 12

    1990

  23. [23]

    Boedecker, O

    J. Boedecker, O. Obst, J. T. Lizier, N. M. Mayer, M. Asada, Theory Biosci. 2012, 131, 205

    2012

  24. [24]

    Bertschinger, T

    N. Bertschinger, T. Natschläger, Neural Computation 2004, 16, 1413

    2004

  25. [25]

    S. M. Popoff, G. Lerosey, M. Fink, A. C. Boccara, S. Gigan, New J. Phys. 2011, 13, 123021

    2011

  26. [26]

    H. Yu, T. R. Hillman, W. Choi, J. O. Lee, M. S. Feld, R. R. Dasari, Y. Park, Phys. Rev. Lett. 2013, 111, 153902

    2013

  27. [27]

    J. Yoon, K. Lee, J. Park, Y. Park, Opt. Express 2015, 23, 10158

    2015

  28. [28]

    X. Lin, Y. Rivenson, N. T. Yardimci, M. Veli, Y. Luo, M. Jarrahi, A. Ozcan, Science 2018, 361, 1004

    2018

  29. [29]

    J. Hu, D. Mengu, D. C. Tzarouchis, B. Edwards, N. Engheta, A. Ozcan, Nat Commun 2024, 15, 1525

    2024

  30. [30]

    Çarpınlıoğlu, U

    B. Çarpınlıoğlu, U. Teğin, Commun Phys 2025, 8, 349

    2025

  31. [31]

    H. Wang, J. Hu, Y. Baek, K. Tsuchiyama, M. Joly, Q. Liu, S. Gigan, Light Sci Appl 2025, 14, 245

    2025

  32. [32]

    Sunada, A

    S. Sunada, A. Uchida, Sci Rep 2019, 9, 19078

    2019

  33. [33]

    H. Cao, R. Chriki, S. Bittner, A. A. Friesem, N. Davidson, Nat Rev Phys 2019, 1, 156

    2019

  34. [34]

    Eliezer, S

    Y. Eliezer, S. Mahler, A. A. Friesem, H. Cao, N. Davidson, Phys. Rev. Lett. 2022, 128, 143901

    2022

  35. [35]

    Teğin, M

    U. Teğin, M. Yıldırım, İ. Oğuz, C. Moser, D. Psaltis, Nat Comput Sci 2021, 1, 542

    2021

  36. [36]

    R. Zhu, J. Hochstetter, A. Loeffler, A. Diaz-Alvarez, T. Nakayama, J. T. Lizier, Z. Kuncic, Sci Rep 2021, 11, 13047

    2021

  37. [37]

    Loeffler, A

    A. Loeffler, A. Diaz-Alvarez, R. Zhu, N. Ganesh, J. M. Shine, T. Nakayama, Z. Kuncic, Sci. Adv. 2023, 9, eadg3289

    2023

  38. [38]

    Mühlnickel, J

    L. Mühlnickel, J. A. Jaurigue, L. C. Jaurigue, K. Lüdge, Commun Phys 2024, 7, 370

    2024

  39. [39]

    W., Saxton, W

    Gerchberg, R. W., Saxton, W. O, Optik 1972, 35

    1972

  40. [40]

    A. Wolf, J. B. Swift, H. L. Swinney, J. A. Vastano, Physica D: Nonlinear Phenomena 1985, 16, 285

    1985

  41. [41]

    Jaeger, Short Term Memory in Echo State Networks, Fraunhofer-Gesellschaft, 2001

    H. Jaeger, Short Term Memory in Echo State Networks, Fraunhofer-Gesellschaft, 2001

    2001