Pith. sign in

REVIEW 3 major objections 5 minor 42 references

A 13,000-parameter model produces calibrated trajectory bands for differential equations it has never seen.

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 · deepseek-v4-flash

2026-08-05 00:12 UTC pith:NW3BDU74

load-bearing objection A genuinely useful hypergraph surrogate for mass-conserving ODEs, with a real overreach in the 'arbitrary parameter distribution' claim that should be fixed before acceptance. the 3 major comments →

arxiv 2608.00852 v1 pith:NW3BDU74 submitted 2026-08-01 cs.LG

HyperODE: Zero-Shot Surrogate for Simulation and Inference of Dynamical Systems

Ajitesh Srivastava This is my paper
classification cs.LG
keywords zero-shot surrogatehypergraph neural networkquantile trajectory predictioncompartmental epidemiological modelsreaction networksamortized Bayesian inferenceuncertainty quantificationneural ODE
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.

This paper claims that a single small neural network can replace the usual per-model surrogate for an entire class of differential equations. The trick is to expose an ODE's structure as a directed hypergraph—each additive term becomes a hyperedge from the states it involves to the state it updates—and to train one shared operator, with an explicit mass-action product, to output trajectory quantiles for any parameter distribution. Once trained on small mass-conserving compartmental and reaction models, the same frozen weights produce calibrated quantile bands for larger systems and unseen families such as SEIR, and the paper shows the operator can also be reversed: a shared encoder maps a noisy observed trajectory to a parameter distribution in milliseconds. If correct, this means a roughly 13,000-parameter model could replace thousands of ODE solves and per-structure retraining for a broad class of models used in epidemiology and chemistry.

Core claim

HyperODE's central claim is that zero-shot transfer across ODE structures is possible if the functional form of the dynamics is separated from the neural architecture. A system's ODE is converted into a directed hypergraph: each state is a node, each additive monomial term is a directed hyperedge with the involved states as sources and the updated state as target, and parameter distributions are attached as quantile-valued edge features. A single shared operator pools source states, multiplies their raw values in a mass-action product, and emits a message scaled by the edge coefficient; integrating the resulting field with a fixed-step solver yields quantile trajectories for every state. Tra

What carries the argument

The central object is a directed hypergraph neural ODE: each additive monomial on the right-hand side of the ODE becomes a hyperedge from its source states to its target state, and every hyperedge is processed by the same small MLP operator shared across the whole system. The operator pools the source hidden states, forms the exact mass-action product of the raw source states, concatenates the product with the hyperedge size and features, and produces a message summed into the target node's derivative. Each node and each edge carries a vector of quantiles rather than a scalar, so the same network represents parameter and trajectory distributions. The explicit product is the load-bearing indu

Load-bearing premise

The transfer claim rests on the trained operator being generic enough that structural changes—new compartments, larger networks, higher-order terms, broken conservation, or time-varying coefficients—fall within what the training distribution of small mass-conserving polynomial systems already taught it.

What would settle it

Take the frozen pretrained operator and evaluate it on a new mass-conserving compartmental family with quadratic monomials but a topology unlike ring-coupled meta-populations, such as a dense random directed reaction graph with cycles, at sizes up to 128 states and parameter distributions it never saw; if the 90% band coverage falls far from 0.90 or the weighted-interval score grows sharply with system size, the zero-shot claim is false.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • A single frozen HyperODE model replaces per-structure surrogates for compartmental and reaction-network ODEs: on held-out SEIR systems up to 257 states and random reaction networks up to 128 states it keeps coverage near 0.90 without retraining.
  • Cross-structure training is not a compromise: the generalist beats a same-architecture specialist trained on each structure by 7–11% in weighted-interval score.
  • For inverse calibration, the shared encoder turns a noisy trajectory into parameter quantiles in 6–13 ms, competitive with per-structure NPE/NRE and with limited-budget MCMC, and the gradient-based inverse is the most accurate method in the comparison.
  • External forcing can be added after training by swapping in a time-varying hypergraph: seasonal SEIR is reproduced to about 2–5% relative error with coverage near 0.91.
  • The product feature can be replaced at inference by any nonnegative driving nonlinearity, letting the same operator integrate non-polynomial incidence functions (saturating, Ricker, Hill) and other ODE families without retraining.

Where Pith is reading between the lines

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

  • An untested extension of the generalized-product result: the frozen operator should be probed on functional families it never saw in any form, such as cubic incidence without polynomialization or rational terms with state-dependent denominators, to see whether calibration holds or only the five reported swaps survive.
  • The paper's two-stage variance argument implies a practical rule for any quantile-surrogate training pipeline: when between-system variance dominates, generate more ODE structures rather than more Monte Carlo rollouts per structure. The author's numbers support this for their class, and the rule should transfer wherever the same loss is used.
  • The encoder is described as the transpose of the forward operator; a natural stress test is to feed it trajectories generated under a different observation model, such as aggregated counts or non-Gaussian noise, to see whether the shared structure still anchors calibration.
  • If the zero-shot claims survive broader holdouts, the same hypergraph representation could be applied to semi-discretized partial differential equations, whose spatial terms are also sums of products of neighboring states; the paper does not attempt this.

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

3 major / 5 minor

Summary. HyperODE is a hypergraph-neural-ODE surrogate for a class of mass-conserving polynomial compartmental/reaction ODEs. Given an ODE structure and a parameter distribution encoded by five quantiles per coefficient, it predicts per-state trajectory quantile bands in one forward pass. The same frozen surrogate is used for inverse calibration, either by gradient descent on parameter quantiles or through a single-pass GRU/message-passing encoder that maps noisy trajectories to parameter quantiles. Training uses SIS/SIR/SI and random reaction networks; evaluation is on held-out SEIR, larger system sizes, higher interaction orders, external forcing, vital dynamics, and selected non-polynomial terms via a product-feature swap. The paper reports WIS and coverage at the MC-oracle level for the forward task and millisecond-scale inverse inference that is competitive with MCMC and per-structure SBI on the tested cases.

Significance. If the central claims hold, this is a meaningful step toward reusable surrogates: a single roughly 13k-parameter model transfers across structures, scales, and interaction orders, and the explicit mass-action product is a clean, well-ablated inductive bias. The paper is stronger than typical surrogate papers because it uses structure-level OOD splits, compares against per-structure specialists, includes an MC-oracle floor and a Monte-Carlo sample-size ANOVA, and releases code. The inverse single-pass encoder, evaluated through the frozen surrogate, is a promising amortized alternative to MCMC for the tested families. The main qualifications are the lossy five-quantile representation of parameter distributions and the manual component of the non-polynomial extensions.

major comments (3)
  1. [Abstract; Quantile (UQ) Forward Surrogate] The abstract claims the method handles 'an arbitrary parameter distribution defined through quantiles.' This is not supported: the input representation is a Q=5 quantile vector per coefficient, and distinct distributions can share the same five quantiles while inducing different trajectory quantiles. Example: A is beta=0.2 with probability 0.5 and beta=0.8 with probability 0.5; B is beta=0.2 with probability 0.5, uniform(0.2,0.8) with probability 0.25, and beta=0.8 with probability 0.25. Both have quantiles (0.2,0.2,0.2,0.8,0.8), but under a nonlinear ODE their Monte-Carlo trajectory quantiles differ. Because HyperODE is deterministic in these edge features, it must emit the same band for both. This is a representational identifiability issue in the core forward machinery, not an OOD generalization issue, and the Appendix's K-sufficiency ANOVA addresses rollout noise rather than input am
  2. [Extension: Generalized Product; Table 8] The zero-shot claims for non-polynomial terms rely on replacing the mass-action product feature with the true nonlinearity g(S,I) and on per-family shifting/normalization and rescaling constants (lambda=91, 811, 3594 in the 'poly. pipeline' column). This requires the user to know the exact functional form and to tune constants; it is not a property of the frozen pretrained model alone. The 'no-prod+poly' baseline also involves hand-specified recasting. These experiments should be presented as proof-of-concept demonstrations with a clear statement of which components are automatic and which require manual specification; as written, 'zero-shot' is overstated.
  3. [Inverse: Noisy Data to Calibration; Single-pass Encoder] The inverse task is described as recovering 'a distribution over the parameters,' but the encoder outputs a deterministic five-quantile vector per parameter, which is a lossy summary rather than a posterior or a full distribution. The paper correctly evaluates the posterior-predictive band rather than parameter recovery in the main experiments, and this is a defensible choice for unidentifiable systems. However, the terminology in the abstract and the method section should be adjusted, and the fact that the single-pass encoder's calibration is evaluated through the same frozen surrogate should be stated explicitly so that any forward-surrogate bias is transparent.
minor comments (5)
  1. [Experiments] Typo: 'We intende to demonstrate' should be 'We intend to demonstrate.' Also, Figure 1's caption and surrounding text refer to k coupled groups, but some notation is used before it is defined; a small notation table would help.
  2. [Method: HyperODE] The product term prod_h is written as a product of the raw state vectors x_s, but it is not explicitly stated that this is an elementwise product over the Q quantile dimensions. Please clarify, since this affects the interpretation of the network.
  3. [Table 1] The generalist-vs-specialist comparison reports WIS only, without standard errors or paired significance tests. Given the small WIS differences (7-11%), adding confidence intervals would strengthen the claim.
  4. [Appendix: Monte-Carlo Sample Size Sufficiency] The two-stage variance formula is a useful contribution, but the text should note that the unbiasedness holds for a fixed predictor; after training on the same rollouts, the empirical loss is not an unbiased estimate of the final model's generalization error. The reported SEs are still useful for a fixed test model.
  5. [Extension: Generalized Product] Figure 10 shows results on oscillators with the statement that the band 'over-covers,' but the reported WIS and coverage numbers for those cases are not given in the main text. Adding quantitative values would make the limitation concrete.

Circularity Check

1 steps flagged

Inverse encoder is evaluated through the same frozen surrogate it was trained against; the forward zero-shot claim is independent.

specific steps
  1. fitted input called prediction [Inverse: Noisy Data to Calibration (Single-pass Encoder) and Appendix Metrics]
    "The encoder is trained through the frozen forward surrogate as a fixed decoder F: it predicts ˆθ=Eη(y), F re-simulates a band, and the loss is the reconstruction WIS against the observation, L(η)=WIS(D( ˆθ), y). ... For the inverse task we report the reconstruction WIS of the posterior-predictive band and coverage based on observed noisy trajectory."

    The inverse encoder is fit by minimizing the WIS between the frozen surrogate's reconstruction D(θ̂) and the observed noisy trajectory y. For the same encoder, the inverse evaluation reports exactly the same reconstruction WIS and coverage of the same posterior-predictive band, generated by the same frozen surrogate. The reported calibration is therefore the training objective evaluated on held-out observations through the fitted simulator, not an independent measure against the true ODE or against MCMC's posterior-predictive band. Systematic surrogate errors can be exploited during training and remain invisible in this closed-loop evaluation, so the inverse 'competitive with existing methods' claim partially reduces to self-consistency. The forward zero-shot surrogate is trained and score

full rationale

The forward surrogate is a standard supervised quantile-regression model: it is trained by minimizing pinball (WIS) loss against Monte-Carlo rollouts of the true ODE and evaluated on out-of-distribution structures. This is genuine generalization, not a derivation loop. The architectural choice of an explicit mass-action product is an inductive bias tested by ablations, not a fitted parameter renamed as a prediction. The extensions (dynamic hypergraphs for forcing, generalized product with known nonlinearities) feed extra known structure into a frozen operator, which is not circular. The main circularity concern is confined to the inverse encoder: both training and evaluation use the same frozen forward surrogate as the decoder and the same reconstruction WIS as the metric, so the inverse calibration claim is partially self-referential. Because the forward claim is independently validated against MC oracles and the paper compares against external benchmarks, the overall circularity score is low.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 4 invented entities

The ledger shows that the contribution is an empirical fit on a carefully chosen training distribution rather than a derivation. The core zero-shot breadth rests on a representational assumption about the hyperedge operator and, for the broadest claims, on per-test tuning constants in the product-swap experiments.

free parameters (5)
  • Forward surrogate network weights (h32-L3) = ~13k parameters
    Trained by pinball loss on K=48 Monte-Carlo rollouts of training ODEs; all generalization claims are properties of these fitted weights, not of a derived operator.
  • Inverse encoder network weights (h128-L2) = ~140k parameters
    Trained through the frozen surrogate with a reconstruction WIS loss and selected by a hyperparameter sweep; inverse performance depends on these fitted weights.
  • K=48 Monte-Carlo rollouts per training system = 48
    Design choice justified by a two-stage variance argument; the target quantiles themselves are estimated from K samples, and different K would change the noise in the training target.
  • Per-family product-swap rescale constants = lambda=91 (SI^0.5), 811 (SI^1.5), 3594 (Hill); none for saturating and Ricker
    Hand-set at inference to rescale timescale or coefficients after polynomialization; these are not trained but are chosen per test family and affect the reported zero-shot WIS and coverage.
  • Architecture and training hyperparameters = width 32, depth 3, lambda_mono=1, learning rate schedule, gradient clipping
    Selected by validation WIS among widths 16/32/64/128 and depths 1/2/3; no sensitivity analysis is reported for most choices.
axioms (6)
  • domain assumption The training distribution of mass-conserving polynomial ODEs is representative of the target class of approximately mass-conserving compartmental and reaction models.
    The zero-shot claim rests on SIS/SIR/SI and random networks generalizing to SEIR, arbitrary random topologies, larger k, and higher order. If the training class is not representative, the OOD results would not generalize. See 'Experiments: Setup' and 'Data Generation'.
  • domain assumption A shared hyperedge operator with an explicit mass-action product can represent the trajectory distribution map for the whole class.
    No approximation theorem is proved; the method relies on the inductive bias and an empirical fit. See Method, hyperedge message equations.
  • domain assumption Quantile-wise composition of separate state and coefficient marginal quantiles can be corrected by the network to produce calibrated trajectory bands.
    Node features are per-state quantile vectors and the product feature multiplies them elementwise, which is not the quantile of the product in general; calibration depends on learned corrections. See Method, product feature definition.
  • standard math Quadratization of non-polynomial ODEs via auxiliary variables is applicable and preserves boundedness and timescales.
    Cites Kerner 1981, Carothers et al. 2005, and Hemery et al. 2020 to justify broad applicability; the paper itself notes unbounded auxiliary states and large coefficients as limitations. See Background and 'Handling non-polynomial terms'.
  • domain assumption The learned operator's locality and time-independence make dynamic hypergraph swaps (forcing) and clamped source or sink nodes (vital dynamics) valid zero-shot.
    The extensions section assumes that replacing a hyperedge product or adding source and sink nodes requires no adaptation of the operator; evidence is limited to a few examples. See 'Extensions' and Appendix 'Extension: Generalized Product'.
  • standard math Two-stage sampling of systems and rollouts gives an unbiased estimate of population WIS with K=48.
    Uses Cochran 1977 two-stage variance; requires finite variance of per-rollout scores and independent sampling of systems. See Appendix 'Monte-Carlo Sample Size Sufficiency'.
invented entities (4)
  • Directed hypergraph representation of an ODE no independent evidence
    purpose: Exposes monomial structure to a shared neural operator.
    An architectural representation, not a physical or falsifiable entity; its value is measured only through downstream predictive performance.
  • Aggregate pseudo-node in random reaction networks no independent evidence
    purpose: Adds an extra state tracking a sum of a subset of derivatives during training.
    Constructed in data generation to provide aggregate signals; no independent empirical handle.
  • Auxiliary clamped source and sink nodes for vital dynamics no independent evidence
    purpose: Models birth and death and non-conservation by adding nodes with clamped values.
    A modeling trick applied at inference; it has no separate falsifiable prediction.
  • Auxiliary state M=1/N for frequency-dependent transmission no independent evidence
    purpose: Polynomializes beta*S*I/N into mass-action form.
    Standard quadratization auxiliary variable; not an observed quantity and no independent evidence is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 15300 in / 20913 out tokens · 241296 ms · 2026-08-05T00:12:18.425977+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of HyperODE: Zero-Shot Surrogate for Simulation and Inference of Dynamical Systems." pith.science (2026). https://pith.science/paper/NW3BDU74

@misc{pith2026260800852,
  author       = {Pith},
  title        = {Pith review of: HyperODE: Zero-Shot Surrogate for Simulation and Inference of Dynamical Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NW3BDU74}},
  note         = {Machine review of arXiv:2608.00852}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Understanding and controlling complex dynamical systems often requires executing thousands of numerical simulations across vast parametric landscapes, which is time-consuming. Machine learning surrogates significantly accelerate simulation by predicting state trajectories across different initializations and parameter values. However, surrogate models are specialized to one simulation model. Modifying the underlying differential equations - e.g., adding a physiological state or altering an epidemiological contact network - renders trained models obsolete and forces computationally expensive retraining from scratch. We introduce HyperODE, a surrogate capable of operating across an entire class of approximately mass-conserving compartmental models without retraining. By mapping the structure of ordinary differential equations (ODEs) into directed hypergraphs, HyperODE decouples the functional form of system interactions from the neural network architecture. HyperODE takes a compartmental model in the form of an ODE with an arbitrary parameter distribution defined through quantiles and transforms it into a hypergraph. It outputs the distribution of the trajectories for all the states in the original ODE in the form of quantiles. We then use this surrogate to build an encoder that takes a noisy trajectory and outputs a distribution over the parameters of the original ODE, thus calibrating the model in a single pass. On families and system sizes never seen in training, HyperODE produces calibrated quantile bands in a single forward pass, with weighted-interval score and coverage on par with specialized surrogates for each structure. For inverse inference, HyperODE produces calibration from noisy state trajectories in a few milliseconds with a single shared encoder, competitive with existing methods. HyperODE extends zero-shot to ODEs that break mass conservation and to external forcing.

Figures

Figures reproduced from arXiv: 2608.00852 by Ajitesh Srivastava.

Figure 1
Figure 1. Figure 1: Simulating a meta-population SIR model over [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of HyperODE forward surrogate. It rep [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Left: Performance of all surrogate models and a sample output. earlier, K need not be large when computing loss across large batches). Trajectories are integrated with a fixed step to tmax=8 (40 steps, dt=0.2). Splits are structure-level out￾of-distribution: only {SIS, SIR, SI} and random networks are seen in training, SIRS is held out for validation, and SEIR for test; the validation/test random networks … view at source ↗
Figure 5
Figure 5. Figure 5: Zero-shot forward generalization to higher interac [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Zero-shot forward scale-generalization (both fam [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 7
Figure 7. Figure 7: Introducing seasonal forcing via a dynamic graph. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: Inverse calibration: runtimes and calibrated bands. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 9
Figure 9. Figure 9: Operator differentiation at h32-L3: interaction order (left) and system scale (right). The choice of explicit product term [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Generalized product across ODE families (zero-shot, one pretrained operator): damped pendulum, Michaelis-Menten, [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Reliability diagrams (empirical vs. nominal cov [PITH_FULL_IMAGE:figures/full_fig_p011_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Median-of-solves vs. solve-at-median on a wide [PITH_FULL_IMAGE:figures/full_fig_p012_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Temporal extrapolation and step-size coarsening [PITH_FULL_IMAGE:figures/full_fig_p012_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Median relative parameter-recovery error vs. sys [PITH_FULL_IMAGE:figures/full_fig_p013_14.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 · 21 canonical work pages · 1 internal anchor

  1. [1]

    1977 , publisher=

    Sampling techniques , author=. 1977 , publisher=

    1977

  2. [2]

    Modeling of Future COVID-19 Cases, Hospitalizations, and Deaths, by Vaccination Rates and Nonpharmaceutical Intervention Scenarios—United States, April--September 2021 , author=. MMWR. Morbidity and Mortality Weekly Report , volume=. 2021 , publisher=

    work page 2021

  3. [3]

    Scientific Reports , volume=

    Statistical inference using GLEaM model with spatial heterogeneity and correlation between regions , author=. Scientific Reports , volume=. 2022 , publisher=

    work page 2022

  4. [4]

    Journal of Mathematical Physics , volume =

    Universal Formats for Nonlinear Ordinary Differential Systems , author =. Journal of Mathematical Physics , volume =

    work page

  5. [5]

    Electronic Journal of Differential Equations , volume =

    Some Properties of Solutions to Polynomial Systems of Differential Equations , author =. Electronic Journal of Differential Equations , volume =

    work page

  6. [6]

    Computational Methods in Systems Biology (CMSB) , year =

    On the Complexity of Quadratization for Polynomial Differential Equations , author =. Computational Methods in Systems Biology (CMSB) , year =

    work page

  7. [7]

    Journal of the Royal Statistical Society: Series B , volume =

    Bayesian Calibration of Computer Models , author =. Journal of the Royal Statistical Society: Series B , volume =

    work page

  8. [8]

    Learning Nonlinear Operators via

    Lu, Lu and Jin, Pengzhan and Pang, Guofei and Zhang, Zhongqiang and Karniadakis, George Em , journal =. Learning Nonlinear Operators via

    work page

  9. [9]

    International Conference on Learning Representations (ICLR) , year =

    Learning Mesh-Based Simulation with Graph Networks , author =. International Conference on Learning Representations (ICLR) , year =

    work page

  10. [10]

    Proceedings of the National Academy of Sciences , volume =

    Machine Learning--Accelerated Computational Fluid Dynamics , author =. Proceedings of the National Academy of Sciences , volume =

  11. [11]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  12. [12]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  13. [13]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    1980

  14. [14]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  15. [15]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

    doi:10.1016/s0020-7373(84)80003-6 1984

  16. [16]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  17. [17]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  18. [18]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

  19. [19]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  20. [20]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

    2023

  21. [21]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

  22. [22]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Neural Ordinary Differential Equations , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  23. [23]

    arXiv preprint arXiv:1911.07532 , year =

    Graph Neural Ordinary Differential Equations , author =. arXiv preprint arXiv:1911.07532 , year =

    Pith/arXiv arXiv 1911

  24. [24]

    International Conference on Learning Representations (ICLR) , year =

    Fourier Neural Operator for Parametric Partial Differential Equations , author =. International Conference on Learning Representations (ICLR) , year =

  25. [25]

    Proceedings of the National Academy of Sciences , volume =

    Discovering Governing Equations from Data by Sparse Identification of Nonlinear Dynamical Systems , author =. Proceedings of the National Academy of Sciences , volume =

    work page

  26. [26]

    arXiv preprint arXiv:2001.04385 , year =

    Universal Differential Equations for Scientific Machine Learning , author =. arXiv preprint arXiv:2001.04385 , year =

    Pith/arXiv arXiv 2001

  27. [27]

    Journal of Computational Physics , volume =

    Physics-Informed Neural Networks: A Deep Learning Framework for Solving Forward and Inverse Problems Involving Nonlinear Partial Differential Equations , author =. Journal of Computational Physics , volume =

  28. [28]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Fast -free Inference of Simulation Models with Bayesian Conditional Density Estimation , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

    work page

  29. [29]

    International Conference on Machine Learning (ICML) , year =

    Automatic Posterior Transformation for Likelihood-Free Inference , author =. International Conference on Machine Learning (ICML) , year =

    work page

  30. [30]

    International Conference on Machine Learning (ICML) , year =

    Likelihood-Free MCMC with Amortized Approximate Ratio Estimators , author =. International Conference on Machine Learning (ICML) , year =

    work page

  31. [31]

    and Mertens, Ulf K

    Radev, Stefan T. and Mertens, Ulf K. and Voss, Andreas and Ardizzone, Lynton and K. IEEE Transactions on Neural Networks and Learning Systems , year =

    work page

  32. [32]

    Foundation Inference Models for Ordinary Differential Equations

    Foundation Inference Models for Ordinary Differential Equations , author =. arXiv preprint arXiv:2602.08733 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  33. [33]

    International Conference on Learning Representations (ICLR) , year =

    d'Ascoli, St. International Conference on Learning Representations (ICLR) , year =

    work page

  34. [34]

    Journal of Open Source Software , volume =

    sbi: A Toolkit for Simulation-Based Inference , author =. Journal of Open Source Software , volume =

    work page

  35. [35]

    AAAI Conference on Artificial Intelligence , year =

    Hypergraph Neural Networks , author =. AAAI Conference on Artificial Intelligence , year =

    work page

  36. [36]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Interaction Networks for Learning about Objects, Relations and Physics , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

    work page

  37. [37]

    International Conference on Machine Learning (ICML) , year =

    Learning to Simulate Complex Physics with Graph Networks , author =. International Conference on Machine Learning (ICML) , year =

    work page

  38. [38]

    Journal of Economic Perspectives , volume =

    Quantile Regression , author =. Journal of Economic Perspectives , volume =

    work page

  39. [39]

    Journal of Machine Learning Research , volume =

    A Tutorial on Conformal Prediction , author =. Journal of Machine Learning Research , volume =

  40. [40]

    PLOS Computational Biology , volume =

    Evaluating Epidemic Forecasts in an Interval Format , author =. PLOS Computational Biology , volume =

    work page

  41. [41]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Hamiltonian Neural Networks , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  42. [42]

    arXiv preprint arXiv:2003.04630 , year =

    Lagrangian Neural Networks , author =. arXiv preprint arXiv:2003.04630 , year =

    Pith/arXiv arXiv 2003