Pith. sign in

REVIEW 4 major objections 6 minor 42 references

Regression can be done by classifying a query against a small library of local patterns and reconstructing the output as a weighted mix of their stored responses.

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-4.5

2026-07-30 13:45 UTC pith:SDOAJWUL

load-bearing objection Clean pattern-library regression with a useful offline cache of responses; works on the toys, but novelty and efficiency claims stay moderate without baselines. the 4 major comments →

arxiv 2607.27196 v1 pith:SDOAJWUL submitted 2026-07-29 cs.LG cs.NAmath.NA

From Classification to Regression: Using a Fruitfly to Solve Equations

Shady E. Ahmed , Panos Stinis This is my paper
classification cs.LG cs.NAmath.NA
keywords classificationregressionbiomimetic algorithmspatternsemergencemachine learningdynamical systemsphysics-informed learning
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 argues that many scientific input–output maps do not need a large global surrogate. Because trajectories and data often revisit a limited set of behaviors, it is enough to store a finite library of representative local patterns and their associated responses. A query is scored for similarity to those patterns, turned into soft weights (via softmax), and the prediction is the weighted combination of the stored responses. The same idea covers continuous-time right-hand sides, discrete flow maps, ordinary data-driven regression, and physics-informed losses, differing only in how the input is embedded and what is stored. Offline work builds the library; online work is only similarity and aggregation, so accuracy, memory, and inference cost trade off through the size of the library. On Lotka–Volterra, Lorenz, simple 1D functions, and 1D Poisson, the method reports competitive accuracy with a compact stored set of patterns.

Core claim

Nonlinear input–output relationships—including dynamical operators, flow maps, and general regression, data-driven or physics-informed—can be replaced by a finite pattern library whose online prediction is softmax-weighted reconstruction of stored pattern responses, giving explicit control of the accuracy–storage–inference trade-off.

What carries the argument

Pattern-library reconstruction: similarities of a query (or its embedding) to stored patterns become softmax activations that form a convex combination of the patterns’ precomputed responses (continuous RHS, discrete flow map, or trainable pattern values), including a fruitfly-style Gaussian “scent” embedding for ordinary regression.

Load-bearing premise

A small finite set of representative patterns, under the chosen similarity and embedding, is expressive enough that convex combinations of their stored responses approximate the true map on the regions that matter, without needing to leave the convex hull of those responses.

What would settle it

On a held-out dynamical regime or 1D target outside the span of the learned patterns, check whether raising pattern count still fails to cut trajectory or test error, or whether recursive prediction drifts because the true response lies outside the convex hull of stored responses.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Online dynamical prediction can drop the global surrogate after training and keep only patterns plus their responses.
  • Accuracy, memory, and inference cost become tunable mainly through pattern count, embedding, and top-R sparsity.
  • The same classification-then-reconstruct pipeline unifies continuous dynamics, discrete maps, data-driven regression, and physics-informed losses.
  • Continual learning can freeze old patterns and add new ones when new regimes appear, without retraining the whole library.
  • Predictions stay inside the convex hull of learned responses, which bounds outputs but also limits extrapolation.

Where Pith is reading between the lines

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

  • If pattern locations and scent variances were trained instead of fixed on a grid with |D|/P, the same library size might resolve sharp features with fewer patterns.
  • The method sits next to classical local and kernel regressors but compresses the reference set to learned prototypes, so scaling studies versus full Nadaraya–Watson or mixture-of-experts would clarify the real online saving.
  • Hull-bounded reconstruction suggests a natural uncertainty flag: large distance from every pattern could mark queries that should not be trusted.

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

4 major / 6 minor

Summary. The paper proposes a unified regression-via-classification framework motivated by fruitfly olfaction: a finite library of representative patterns stores associated responses, and a query is mapped to a prediction by softmax-weighted reconstruction of those responses (Eqs. 7–9, 16, 24–25). The same pipeline is applied to continuous-time RHS and discrete flow-map learning for dynamical systems, and—via a Gaussian “scent” embedding—to data-driven and physics-informed 1D regression. An offline–online split is emphasized so that online inference reduces to similarity evaluation and aggregation. Demonstrations include Lotka–Volterra (known and unknown dynamics), Lorenz recursive flow maps with a continual-learning augmentation, and 1D parabola/sine/Poisson examples with relative errors typically ~0.1–3% (up to ~10% on a non-smooth forcing).

Significance. If the approach scales beyond the reported low-dimensional demos, it would offer a compact, controllable alternative to global surrogates for scientific regression and dynamical prediction, with an explicit accuracy–storage–inference knob via library size P and optional top-R sparsity. Strengths include a clean, internally consistent reconstruction formula; a genuine offline–online decomposition; and a unified treatment of operator learning, flow maps, supervised regression, and physics-informed least squares. The continual pattern-augmentation idea for Lorenz is a useful practical contribution. The work does not ship machine-checked proofs or public code in the manuscript, and the efficiency advantage over global models is argued structurally rather than measured; significance therefore hinges on whether modest libraries plus the chosen similarities remain expressive on scientifically relevant regions—an assumption the authors themselves flag.

major comments (4)
  1. [Abstract; §1; §4.1.2; §6] Abstract, §1, and §6 claim reduced computational/memory demand and explicit accuracy–storage–inference control relative to global surrogates, yet the manuscript reports no FLOPs, wall-clock inference, peak memory, or parameter-count comparisons against the neural emulators used offline, nor against standard local/kernel baselines (Nadaraya–Watson, RBF, partition-of-unity, piecewise-affine). For Lotka–Volterra (§4.1.2) a 3×128 network is trained for ~500K iterations then discarded; only trajectory error is shown. Without a quantitative baseline table, the central efficiency claim remains unsupported even if the reconstruction formulas are correct.
  2. [§3.1–3.2; §4.1–4.2] §3.1–3.2 introduce the pattern library P={v_k} and losses L_RHS / L_traj, but the optimization of the pattern locations v_k themselves is underspecified: initialization, whether v_k are free parameters or selected from data, joint vs staged training with f_θ/M_θ, and any regularization against redundant patterns are not stated. Figures 1–3 and 5–7 cannot be reproduced from the text alone, and the claimed compactness (nP parameters) depends on this procedure. A precise algorithm box or pseudocode for offline pattern identification is needed.
  3. [§3.1 (after Eq. 9); §3.3; §5; §6] After Eq. (9) predictions are confined to the convex hull of stored responses; §3.3 further fixes scent variances to |D|/P and uses equidistant pattern sites in 1D only. All dynamical demos are 2–3D and regression demos are scalar 1D. This is consistent but load-bearing for the generality claim (“nonlinear input–output relationships,” “scientific data”). The manuscript should either (i) demonstrate at least one higher-dimensional or non-convex-response regime where the hull limitation is stress-tested, or (ii) clearly scope the claims to settings where local responses are well approximated by convex combinations of a small library, with failure modes discussed.
  4. [§4.2; Fig. 5–7] Lorenz results (§4.2, Fig. 5) report ~1% average ℓ2 on t∈[0,10] recursive rollouts and a continual-learning gain after reinitialization at t=8 (Fig. 7). For a chaotic system this metric alone is weak: there is no reporting in Lyapunov-time units, attractor statistics (e.g., climate error), ensemble of ICs beyond “10 trajectories,” or comparison to simply retaining the trained flow-map network. The reinitialization step also removes the hardest part of long-horizon chaos, so the continual-learning claim should be restated more carefully and supplemented with chaos-appropriate diagnostics.
minor comments (6)
  1. [§4.1; affiliations; figures] Typos and naming inconsistencies: “Lotka-V oletrra” / “Lotka-Voletrra” (Fig. 1 caption), “W A” in affiliations, double commas in parameter lists (§4.1), and “Fruitfly” vs “fruit fly” / “fruitfly” mixed throughout.
  2. [Fig. 2–3, 6, 8–9] Figure 8–9 captions begin with a stray period (“Figure 8: . Results…”). Several figures lack axis labels or colorbar legends sufficient to read pattern-index partitions (Fig. 2–3, 6).
  3. [§3.1; Eq. 10] Softmax temperature τ is introduced as possibly learnable or fixed to 1/√n (§3.1) but never reported for the experiments; likewise top-R / ε sparsity is defined (Eq. 10) but unused in results.
  4. [§1; §3.3] Related-work positioning vs kernel/Nadaraya–Watson and mixture-of-experts is present but brief; a short explicit contrast table (what is stored, what is evaluated online, dependence on N vs P) would help readers.
  5. [§5.2; Eq. 32] Physics-informed Poisson (§5.2): how second derivatives of wk are obtained (analytic vs autodiff) and enforcement of Dirichlet BCs are not stated; BC treatment affects the least-squares problem in Eq. (32).
  6. [Title page] arXiv id / date line says “July 30, 2026” which appears to be a placeholder; fix metadata before journal submission.

Circularity Check

0 steps flagged

Method-plus-demo paper; pattern library is fit by design and evaluated on held-out/integrated quantities—no derivation that redefines the target as its own input.

full rationale

The central construction (Eqs. 7–9, 16, 24–25) is an explicit surrogate: learn a finite pattern library offline, then form online predictions as softmax-weighted convex combinations of stored responses. Pattern values (and optional emulator weights) are optimized against data residuals, trajectory rollouts, or physics residuals by design; reported errors are on integrated trajectories, test points, or PDE residuals, not tautological restatements of the fit. The convex-hull limitation after Eq. 9 is an expressivity bound the authors state openly, not a hidden circular reduction. Self-citations ([36], [41], [42]) are peripheral related-work pointers (multifidelity/continual learning, chaos-aided classification) and do not underwrite uniqueness or force the main claims. Distinctions from kernels, Nadaraya–Watson, local regression, and MoE are argued rather than smuggled. No step reduces a claimed prediction to its fitted input by construction. Score 1 only for ordinary non-load-bearing author self-citation.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 2 invented entities

Load-bearing content is methodological: softmax-weighted reconstruction from a finite pattern set; similarity via inner products or Gaussian scent features; optional NN emulators only offline; hand-chosen pattern counts, temperatures, variances, and sampling grids. No new physical law is postulated. Claims rest on standard ODE/regression setup plus the modeling assumption that few recurring local behaviors suffice and that convex mixtures of stored responses approximate the map.

free parameters (7)
  • Number of patterns P = task-dependent (e.g. 16–32 dynamics; 10–100 regression)
    Primary capacity knob; swept in LV (various P), fixed 32/16/extra-16 in unknown-dynamics and Lorenz cases, and 10–100 in regression/Poisson examples. Directly controls accuracy–storage trade-off claimed in the abstract.
  • Softmax temperature τ = not systematically reported per experiment
    Controls sharpness of pattern activation (Eq. 8); described as learnable or fixed (e.g. 1/√n). Affects which patterns contribute.
  • Scent variances σ_P² and σ_Q² = |D|/P and |D|/n
    User-prescribed Gaussian widths in the fruitfly regressor; set to |D|/P and |D|/n in all reported 1D examples; authors note this choice is counterintuitive vs ballistic scaling and critical to accuracy (§3.3, §6).
  • Scent sampling dimension n and sample locations z_i = n=P in reported 1D runs
    Dimension of scent vectors and the grid where Gaussians are evaluated; typically set equal to P with domain samples. Changes the similarity geometry.
  • Trainable pattern responses ũ_Pk or f_θ(v_k) / M_θ(v_k) = fit per task (least squares or NN training)
    Core fitted quantities: pattern output values in regression/Poisson, or emulator outputs at patterns in unknown dynamics. Fit by least squares, trajectory losses, or physics residuals.
  • Neural emulator architecture and training budget = e.g. 3 hidden layers × 128; ~500K iters (LV)
    When dynamics are unknown, a feedforward net (e.g. 3×128 for LV flow map; similar for Lorenz) and long in-the-loop optimization define the cached responses. Not needed online but shapes the library.
  • Top-R / threshold ε sparsity (optional) = optional; not central in reported figures
    Optional inference sparsification (Eq. 10); another accuracy–cost knob if used.
axioms (5)
  • standard math Standard autonomous ODE / discrete flow-map and supervised regression problem statements (Eqs. 1–3, 15, 18).
    Background problem class assumed throughout §2.
  • domain assumption Scientific trajectories/data occupy limited recurring regions of input space, so a finite pattern library can represent the map locally.
    Stated in abstract and §1 as the reason global surrogates can be replaced; without it the finite-P approach fails.
  • ad hoc to paper Similarity in the chosen embedding (inner product on states, or Gaussian scent overlap) ranks dynamically/functionally relevant patterns.
    Method definition in §3.1 and §3.3; not derived from a theorem that this similarity is optimal.
  • ad hoc to paper Softmax weights yield a useful convex combination of stored responses (prediction in conv(hull) of pattern responses).
    Eqs. 8–9 and explicit convex-hull limitation in §3.1; extrapolation outside represented behaviors is disclaimed rather than solved.
  • domain assumption For physics-informed Poisson, second derivatives of softmax weights times pattern values form a valid residual minimizer under Dirichlet conditions.
    §5.2 loss (Eq. 32); standard PINN-style residual assumption with this particular ansatz.
invented entities (2)
  • Pattern library P = {v_k} with stored responses as computational primitives for regression-via-classification no independent evidence
    purpose: Replace a global surrogate by classify-then-reconstruct over representative local behaviors.
    Central construct of §3; related to prototypes/clustering but used specifically as prediction atoms, not data compression alone.
  • Fruitfly regressor (Gaussian scent embeddings for query and pattern points, sampled to vectors, softmax mix of trainable ũ_Pk) no independent evidence
    purpose: Extend the pattern framework to non-state regression and physics-informed 1D problems via a biomimetic embedding.
    Defined in §3.3 Eqs. 20–25; motivated by fruit-fly olfaction literature but the specific regressor is introduced here.

pith-pipeline@v1.2.0-daily-grok45 · 18165 in / 4164 out tokens · 91621 ms · 2026-07-30T13:45:54.666104+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of From Classification to Regression: Using a Fruitfly to Solve Equations." pith.science (2026). https://pith.science/paper/SDOAJWUL

@misc{pith2026260727196,
  author       = {Pith},
  title        = {Pith review of: From Classification to Regression: Using a Fruitfly to Solve Equations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SDOAJWUL}},
  note         = {Machine review of arXiv:2607.27196}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We present a novel approach to regression tasks using classification which is motivated by the mechanism used by fruitflies to sense their environment. Specifically, we formulate a general framework for learning nonlinear input-output relationships by replacing complex global surrogate models with a finite library of representative local patterns. Since scientific data often occupy limited and recurring regions of the input space, we generate predictions by measuring similarities between a query and stored patterns, then combining their associated responses through weighted reconstruction. We apply this approach to nonlinear dynamical systems, data-driven regression, and physics-informed learning using suitable embeddings and similarity measures. For dynamical systems, our offline-online workflow extracts patterns from data or governing equations during the offline phase, while online prediction requires only similarity evaluation and response aggregation. This structure helps us reduce computational and memory demands while providing explicit control over the trade-off among accuracy, storage, and inference cost.

Figures

Figures reproduced from arXiv: 2607.27196 by Panos Stinis, Shady E. Ahmed.

Figure 1
Figure 1. Figure 1: Prediction of the Lotka-Voletrra dynamics using different numbers of patterns along with the associated [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Pattern dominance regions for the Lotka-Volterra system. Each point in the state space is colored according to [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Assigned weight for each pattern in different regions of the state space for [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Prediction of the Lotka–Volterra system without access to the governing equations. A surrogate model is used [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Long-term prediction of the Lorenz system trajectory from a test initial condition using [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Temporal evolution of pattern weights during a Lorenz system trajectory. Different patterns become active at [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Continual learning through pattern augmentation. After training on data from [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: . Results for function regression. (Left) Parabola. (Middle) A random combination of 10 sines with [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: . Results for 1D Poisson with different forcing functions. (Left) Parabola. (Middle) A random combination of [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] 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

42 extracted references · 6 linked inside Pith

  1. [1]

    Reduced-order modeling: new approaches for computational physics

    David J Lucia, Philip S Beran, and Walter A Silva. Reduced-order modeling: new approaches for computational physics. Progress in Aerospace Sciences, 40(1-2):51–117, 2004

    2004

  2. [2]

    Model reduction and approximation: theory and algorithms

    Peter Benner, Mario Ohlberger, Albert Cohen, and Karen Willcox. Model reduction and approximation: theory and algorithms. SIAM, 2017

    2017

  3. [3]

    Model reduction for flow analysis and control

    Clarence W Rowley and Scott TM Dawson. Model reduction for flow analysis and control. Annual Review of Fluid Mechanics, 49:387–417, 2017

    2017

  4. [4]

    Kernel methods in machine learning

    Thomas Hofmann, Bernhard Schölkopf, and Alexander J Smola. Kernel methods in machine learning. 2008

    2008

  5. [5]

    Kernel methods for pattern analysis

    John Shawe-Taylor and Nello Cristianini. Kernel methods for pattern analysis. Cambridge University Press, 2004

    2004

  6. [6]

    Gaussian processes for regression

    Christopher Williams and Carl Rasmussen. Gaussian processes for regression. Advances in neural information processing systems, 8, 1995

    1995

  7. [7]

    A tutorial on gaussian process regression: Modelling, exploring, and exploiting functions

    Eric Schulz, Maarten Speekenbrink, and Andreas Krause. A tutorial on gaussian process regression: Modelling, exploring, and exploiting functions. Journal of Mathematical Psychology, 85:1–16, 2018

    2018

  8. [8]

    An intuitive tutorial to gaussian process regression

    Jie Wang. An intuitive tutorial to gaussian process regression. Computing in Science & Engineering, 25(4):4–11, 2023

    2023

  9. [9]

    Radial basis functions

    Martin Dietrich Buhmann. Radial basis functions. Acta Numerica, 9:1–38, 2000

    2000

  10. [10]

    A radial basis function method for the shallow water equations on a sphere

    Natasha Flyer and Grady B Wright. A radial basis function method for the shallow water equations on a sphere. Proceedings: Mathematical, Physical and Engineering Sciences, pages 1949–1976, 2009

    1949

  11. [11]

    Deep learning: Foundations and concepts

    Christopher M Bishop and Hugh Bishop. Deep learning: Foundations and concepts. Springer Nature, 2023

    2023

  12. [12]

    Understanding deep learning

    Simon JD Prince. Understanding deep learning. MIT press, 2023

    2023

  13. [13]

    Deep learning with Python

    Francois Chollet and François Chollet. Deep learning with Python. simon and schuster, 2021

    2021

  14. [14]

    Testing the manifold hypothesis

    Charles Fefferman, Sanjoy Mitter, and Hariharan Narayanan. Testing the manifold hypothesis. arXiv preprint arXiv:1310.0425, 2013

    Pith/arXiv arXiv 2013

  15. [15]

    Geometry from a time series

    Norman H Packard, James P Crutchfield, J Doyne Farmer, and Robert S Shaw. Geometry from a time series. Physical Review Letters, 45(9):712, 1980

    1980

  16. [16]

    Locally weighted regression: an approach to regression analysis by local fitting

    William S Cleveland and Susan J Devlin. Locally weighted regression: an approach to regression analysis by local fitting. Journal of the American Statistical Association, 83(403):596–610, 1988

    1988

  17. [17]

    The approximation power of moving least-squares

    David Levin. The approximation power of moving least-squares. Mathematics of Computation, 67(224):1517– 1531, 1998

    1998

  18. [18]

    Analysis of moving least squares approximation revisited

    Davoud Mirzaei. Analysis of moving least squares approximation revisited. Journal of Computational and Applied Mathematics, 282:237–250, 2015

    2015

  19. [19]

    The partition of unity finite element method: basic theory and applications

    Jens M Melenk and Ivo Babuška. The partition of unity finite element method: basic theory and applications. Computer Methods in Applied Mechanics and Engineering, 139(1-4):289–314, 1996

    1996

  20. [20]

    The partition of unity method

    Ivo Babuška and Jens M Melenk. The partition of unity method. International Journal for Numerical Methods in Engineering, 40(4):727–758, 1997. 15 Using a Fruitfly to Solve EquationsA PREPRINT

    1997

  21. [21]

    Probabilistic partition of unity networks for high-dimensional regression problems

    Tiffany Fan, Nathaniel Trask, Marta D’Elia, and Eric Darve. Probabilistic partition of unity networks for high-dimensional regression problems. International Journal for Numerical Methods in Engineering, 124(10): 2215–2236, 2023

    2023

  22. [22]

    A clustering technique for the identification of piecewise affine systems

    Giancarlo Ferrari-Trecate, Marco Muselli, Diego Liberati, and Manfred Morari. A clustering technique for the identification of piecewise affine systems. Automatica, 39(2):205–217, 2003

    2003

  23. [23]

    Piecewise affine regression via recursive multiple least squares and multicategory discrimination

    Valentina Breschi, Dario Piga, and Alberto Bemporad. Piecewise affine regression via recursive multiple least squares and multicategory discrimination. Automatica, 73:155–162, 2016

    2016

  24. [24]

    A piecewise linear regression and classification algorithm with application to learning and model predictive control of hybrid systems

    Alberto Bemporad. A piecewise linear regression and classification algorithm with application to learning and model predictive control of hybrid systems. IEEE Transactions on Automatic Control, 68(6):3194–3209, 2022

    2022

  25. [25]

    Nonlinear model order reduction based on local reduced- order bases

    David Amsallem, Matthew J Zahr, and Charbel Farhat. Nonlinear model order reduction based on local reduced- order bases. International Journal for Numerical Methods in Engineering, 92(10):891–916, 2012

    2012

  26. [26]

    Breaking the kolmogorov barrier in model reduction of fluid flows

    Shady E Ahmed and Omer San. Breaking the kolmogorov barrier in model reduction of fluid flows. Fluids, 5(1): 26, 2020

    2020

  27. [27]

    Can a fruit fly learn word embeddings? arXiv preprint arXiv:2101.06887, 2021

    Yuchen Liang, Chaitanya K Ryali, Benjamin Hoover, Leopold Grinberg, Saket Navlakha, Mohammed J Zaki, and Dmitry Krotov. Can a fruit fly learn word embeddings? arXiv preprint arXiv:2101.06887, 2021

    Pith/arXiv arXiv 2021

  28. [28]

    Embedology

    Tim Sauer, James A Yorke, and Martin Casdagli. Embedology. Journal of Statistical Physics, 65(3):579–616, 1991

    1991

  29. [29]

    Adaptive mixtures of local experts

    Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural Computation, 3(1):79–87, 1991

    1991

  30. [30]

    Out- rageously large neural networks: The sparsely-gated mixture-of-experts layer

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Out- rageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017

    Pith/arXiv arXiv 2017

  31. [31]

    Ensemble and mixture-of-experts DeepONets for operator learning

    Ramansh Sharma and Varun Shankar. Ensemble and mixture-of-experts DeepONets for operator learning. arXiv preprint arXiv:2405.11907, 2024

    Pith/arXiv arXiv 2024

  32. [32]

    Regression by classification

    Luís Torgo and Joao Gama. Regression by classification. In Brazilian symposium on artificial intelligence, pages 51–60. Springer, 1996

    1996

  33. [33]

    Regression as classification

    Raied Salman and V ojislav Kecman. Regression as classification. In 2012 Proceedings of IEEE Southeastcon, pages 1–6. IEEE, 2012

    2012

  34. [34]

    Cluster-based reduced-order modelling of a mixing layer

    Eurika Kaiser, Bernd R Noack, Laurent Cordier, Andreas Spohn, Marc Segond, Markus Abel, Guillaume Daviller, Jan Östh, Siniša Krajnovi´c, and Robert K Niven. Cluster-based reduced-order modelling of a mixing layer. arXiv preprint arXiv:1309.0524, 2013

    Pith/arXiv arXiv 2013

  35. [35]

    Cluster-based network modeling—from snapshots to complex dynamical systems

    Daniel Fernex, Bernd R Noack, and Richard Semaan. Cluster-based network modeling—from snapshots to complex dynamical systems. Science Advances, 7(25):eabf5006, 2021

    2021

  36. [36]

    A multifidelity deep operator network approach to closure for multiscale systems

    Shady E Ahmed and Panos Stinis. A multifidelity deep operator network approach to closure for multiscale systems. Computer Methods in Applied Mechanics and Engineering, 414:116161, 2023

    2023

  37. [37]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations

    Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686–707, 2019. doi:10.1016/j.jcp.2018.10.045

    doi:10.1016/j.jcp.2018.10.045 2019

  38. [38]

    Weighted Nadaraya–Watson regression estimation

    Zongwu Cai. Weighted Nadaraya–Watson regression estimation. Statistics & Probability Letters, 51(3):307–318, 2001

    2001

  39. [39]

    A least squares radial basis function partition of unity method for solving pdes

    Elisabeth Larsson, Victor Shcherbakov, and Alfa Heryudono. A least squares radial basis function partition of unity method for solving pdes. SIAM Journal on Scientific Computing, 39(6):A2538–A2563, 2017

    2017

  40. [40]

    A comprehensive survey of continual learning: Theory, method and application

    Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8):5362–5383, 2024

    2024

  41. [41]

    A multifidelity approach to continual learning for physical systems

    Amanda Howard, Yucheng Fu, and Panos Stinis. A multifidelity approach to continual learning for physical systems. Machine Learning: Science and Technology, 5(2):025042, 2024

    2024

  42. [42]

    Enhancing classification accuracy through chaos

    Panos Stinis. Enhancing classification accuracy through chaos. arXiv preprint arXiv:2603.15299, 2026. 16

    Pith/arXiv arXiv 2026