Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Control of Overfitting with Physics

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims that the Eyring formula of chemical kinetics explains why stochastic gradient Langevin dynamics and GAN training suppress narrow minima and thereby control overfitting.

desk verdict A genuinely nice SGLD/Eyring analogy under an under-supported GAN/predator-prey analogy; the GAN conclusion needs real evidence or a derivation from the minimax dynamics. read the letter →

arxiv 2412.10716 v1 pith:MPZKC47P submitted 2024-12-14 cs.LG cs.AIstat.ML

Sergei V. Kozyrev , Ilya A Lopatin , Alexander N Pechen This is my paper · ORCID
classification cs.LGcs.AIstat.ML
keywords overfittingcontrolstochasticgradientLangevindynamicsEyringformulafreeenergygenerativeadversarialnetworkspredator-preymodelalgorithmicstabilitywideminima
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish a physical mechanism behind overfitting control in machine learning. It argues that under stochastic gradient Langevin dynamics, the Eyring formula from chemical kinetics makes the escape rate out of a potential well proportional to $e^{-\beta(F_1-F_0)}$, where wider wells have higher entropy and therefore lower free energy, so they capture the optimizer more effectively. Because algorithmic stability ties wide minima to good generalization, noise injection becomes a free-energy optimization that suppresses overfitting. For generative adversarial networks, the paper proposes a predator–prey model in which the generator pushes the discriminator out of narrow likelihood maxima, and it gives simulations on a two-well potential and a small dataset that show the predicted wide-maximum selection.

What carries the argument

The load-bearing identities are the Eyring rate formula $r \propto e^{-\beta(F_1-F_0)}$ with free energy $F = E - \beta^{-1}S$ (which turns well width into capture probability), and the Fokker–Planck equation (5) whose stationary solution is the Gibbs measure $e^{-\beta f}$. These are applied to the SGLD dynamics (3)–(4). For the GAN, the machinery is the two-body SDE system (9)–(10) built on the minimax functional (8); the hand-crafted interaction potential (18), containing a short-range Yukawa repulsion and a mid-range attraction, is what implements the wide-maximum selection in the simulations and in the proposed design rules.

What would settle it

On a synthetic two-well potential with equal barrier heights but unequal well widths, run SGLD for many temperatures and compare the fraction of trajectories ending in each well against the Eyring prediction; a systematic mismatch, such as narrow-well capture increasing with temperature when free energy orders them oppositely, would falsify the entropy-selection claim. For the GAN part, train a GAN on a toy data distribution engineered to have only a narrow likelihood peak: if the interaction potential (18) does not push the discriminator out of that peak at the predicted parameter scale, the predator–prey mechanism is not faithful to actual GAN training.

Watch

Extended reading notes

Core claim

The paper's central claim is that overfitting control in SGLD and GAN training follows quantitatively from the Eyring formula: the rate at which a stochastic optimizer escapes a potential well is $r \propto e^{-\beta(F_1-F_0)}$, where $F_1$ is the free energy of the saddle between wells and $F_0$ that of the initial well, with $F = E - \beta^{-1}S$. Wider wells have higher entropy $S$, lower free energy, and thus longer residence times; therefore stochastic gradient Langevin dynamics concentrates on wide minima of the empirical risk, which the algorithmic stability framework identifies with low overfitting. For the GAN, the paper extends this to a two-body system of stochastic differential equations (9)–(10) with the minimax potential (8), interpreting the discriminator as prey and the generator as predator. The predator–prey interaction adds a repulsive force that pushes the discriminator out of narrow likelihood maxima; a branching random process generalization, in which discriminators and generators replicate and die, is argued to strengthen this suppression. Numerical simulations on a double-well objective and a wine recognition dataset demonstrate the predicted regimes of escape from narrow wells and oscillation in wide wells.

Load-bearing premise

The loss landscapes of real neural networks must have well-defined minima and saddles for the Eyring formula to apply; the paper explicitly states that for landscapes without such structure, the approach may not work.

Editorial extensions

If this is right

  • Noise injection in SGLD can be treated as free-energy optimization: learning rate and temperature jointly control which width of minima survive, giving a quantitative overfitting dial.
  • GAN training should show measurably stronger overfitting reduction than plain discriminative training on the same data, because the predator–prey term adds an extra suppression of narrow likelihood maxima.
  • The branching random process version of the GAN predicts that increasing replication rates strengthens wide-maximum selection, so population-based GAN variants would be expected to generalize even better.
  • The Eyring formula gives a concrete, testable prediction for the ratio of capture probabilities into two minima as a function of barrier height, well width, and temperature, which can be checked in controlled stochastic-optimization experiments.
  • If the mechanism holds, explicit flat-minima regularizers should be interchangeable with simply raising the noise level in SGLD, a prediction practitioners can test.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If entropy-based selection is the true cause of overfitting control, then many seemingly unrelated heuristics—dropout, label smoothing, early stopping—may be unified as different ways of raising the effective temperature of the loss landscape.
  • The predator–prey analogy suggests a design principle beyond the paper's specific potential (18): by choosing the interaction force's characteristic length scales, one could in principle program which widths of likelihood maxima survive, a knob that toy simulations could test before scaling to real generators.
  • The branching random process model makes population-level predictions—for instance, that narrow maxima become extinction traps for the discriminator population—which could be tested by tracking per-mode particle counts during GAN training, an experiment the paper does not run.
  • The same free-energy argument might transfer to other stochastic optimizers whose dynamics converge to a Gibbs measure, such as certain randomized coordinate descent or Gibbs-sampling-based learners, giving a general theory of why noise helps generalization.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes physics- and biology-based analogies to explain overfitting control in machine learning. For stochastic gradient Langevin dynamics (SGLD), it invokes the Eyring formula of kinetic theory to argue that wider minima, having lower free energy, are selected and hence algorithmic stability is improved. For generative adversarial networks (GANs), it draws an analogy between GAN training and a predator–prey system, claiming that this dynamics implements selection of wide likelihood maxima and reduces overfitting. The paper also introduces a branching random process generalization of GANs. Numerical simulations include a 2D SGLD experiment on a two-well objective and a predator–prey ODE simulation on the same type of objective, plus a small Wine dataset regression experiment.

Significance. If substantiated, the SGLD/Eyring connection would provide a principled, externally grounded explanation for the tendency of stochastic gradient methods to find flat minima and generalize, complementing algorithmic stability theory. The paper’s honest acknowledgment of the Eyring formula’s landscape assumptions and its reference to rigorous kinetic-theory results are strengths. However, the GAN/predator–prey contribution is currently an analogy rather than an established mechanism: the interaction potential used in simulations is hand-crafted and not derived from the GAN minimax structure, and no actual GAN experiment is performed. The branching-process generalization is a proposal without analysis. The paper is likely to stimulate intuition but, as it stands, it does not fully support the stated GAN overfitting conclusion.

major comments (3)
  1. [§3.2 and §5.3] The central GAN claim that 'the GAN model implements the selection of wide likelihood maxima, which reduces overfitting' is not derived from the GAN minimax SDE system (9)–(10). The predator–prey system (16)–(18) is introduced ad hoc: the prey maximizes a scalar objective L plus a hand-chosen force V(x−y), while the predator chases at constant speed; the interaction in real GAN training is mediated by the KL-type term in (8) on the space of distributions, not by the distance-dependent vector potential in (18). No argument shows that (18) realizes the effective dynamics of (8). Consequently, the simulations in Section 5.3 demonstrate behavior of the chosen ODE, not of GAN training.
  2. [§5.4] The Wine dataset experiment applies the same predator–prey ODE (16)–(18) to quadratic regression, not the GAN SDE system (9)–(10). Therefore Table 1 does not provide evidence about GAN training. To support the GAN overfitting conclusion, the authors would need either to derive (18) from the structure of (8) or to simulate the actual system (9)–(10) on a controlled two-well discriminator objective and show that narrow likelihood maxima are suppressed relative to the SGLD baseline.
  3. [§2.2] The SGLD/Eyring argument assumes the empirical risk landscape has well-defined minima and saddles so that the Eyring formula applies. The manuscript explicitly acknowledges that 'for landscapes which might not exhibit clear minima and saddles between them, the proposed approach based on the Eyring formula may not work.' While this acknowledgment is commendable, it means the claim that SGLD 'allows to control overfitting' is conditional: it holds only for landscapes of the required topological type, and the paper provides no evidence that real neural loss landscapes satisfy this condition. The authors should state this condition as a hypothesis and discuss evidence from the loss-landscape literature, rather than presenting the conclusion as unconditional.
minor comments (5)
  1. [Eq. (16)] The predator–prey equation is dimensionally inconsistent as written: L(x) is a scalar objective function in Eq. (15), while dx/dt and V(x−y) are vectors. The authors presumably mean ∇L(x) (or a gradient-like term); this should be corrected to make the model well-defined.
  2. [§5.3] The functions V and W in Eqs. (17)–(18) are called 'potentials' but are vector functions (forces). Using 'potential' is misleading; the authors should consistently call them interaction forces or specify the scalar potential from which V is derived.
  3. [§5.2] The SGLD simulation uses an annealing schedule with variance proportional to (1+k)^{-1/2}, so the process is not run at a fixed temperature T. The reader should be told how this relates to the stationary Gibbs distribution at temperature T referenced in the Eyring argument.
  4. [§5.4] Table 1 reports point estimates without error bars or repeated-seed statistics. The text says 'similar results in general' but gives no quantitative support; a small table with mean and standard deviation over several splittings would be more convincing.
  5. [§4] The branching random process generalization is introduced with proposed replication and death rates, but no analysis or simulation is provided. If this section is meant to be a substantive contribution, the claim that overfitting control is 'more pronounced' than for standard GANs needs support; otherwise it should be framed as a speculative research direction.

Circularity Check

1 steps flagged · score 6.0 of 10

The SGLD/Eyring analysis is independent, but the GAN wide-minimum claim is partly circular: the simulation's interaction potential is constructed to produce the very push-out and oscillation behavior the paper reports.

  1. self definitional [Section 5.3, Eq. (18) and the bullet conditions on V preceding it]
    "A key point in our analysis is to find suitable conditions on the vector function V (x(t) − y(t))... At short distances, d < σmin... we assume ∥V ∥ ≫ 1 to allow a predator to push the prey out of the well. At intermediate distances... we assume ∥V ∥ ∼ ∥W ∥. This condition is introduced to have oscilations in sufficiently wide wells."

    The vector interaction V in Eq. (18) is selected so that, by construction, short predator–prey distances expel the prey from narrow wells and intermediate distances sustain oscillations in wide wells. The simulation then reports exactly these two regimes ('pushing out of the narrower well and oscillations in a wider well'). Since the paper states Eq. (16) is 'a more general dynamical system than the system defined by Equations (9) and (10)', the observed wide-maximum selection is not a prediction of GAN dynamics; it is an assumption baked into the toy model. The GAN overfitting conclusion therefore rests on a model whose defining conditions already contain the claimed outcome.

full rationale

The SGLD/Eyring portion (Sections 2.1–2.2) is not circular: it imports the Eyring formula from chemical kinetics [24], defines free energy explicitly, and the 2D SGLD simulation confirms the predicted preference for wider wells as temperature increases. The stated limitation that Eyring reasoning fails for landscapes without clear minima and saddles is an honest scope restriction, not a circular step. The GAN portion is different. The minimax SDEs (9)–(10) follow from the GAN objective, but the paper's central claim that the GAN 'implements the selection of wide likelihood maxima' is supported by the predator–prey model of Eqs. (16)–(18), whose interaction V is explicitly chosen to push prey out of narrow wells and oscillate in wide ones. The simulation is therefore a demonstration of the model's own defining conditions rather than an independent test of the GAN mechanism, making the GAN overfitting reduction claim partially circular or at least established only by construction. The Wine benchmark does test the predator–prey ODE externally, but it does not test the GAN SDE system, so it does not remove the circularity in transferring the toy-model behavior to GANs. Self-citations such as [8] and [34] are pointers or background, not load-bearing for the main derivation. Overall, the SGLD result is independent; only the GAN selection claim reduces by construction, hence score 6.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The SGLD part uses no fitted parameters beyond the simulation's temperature and well geometry. The GAN predator-prey part introduces a hand-tuned interaction potential with four to five parameters that are chosen to reproduce the desired qualitative regimes, which is a form of fitting to the phenomenon being explained. The Eyring formula and the flat-minima-to-generalization link are imported as axioms from prior literature.

free parameters (4)
  • Temperature T in SGLD simulation = scanned from 0 to 0.8
    Controls noise amplitude in SGLD; increasing T shifts selection toward the wider well (Figure 2).
  • Objective well parameters (sigma1, sigma2, c1, c2, q_j) = sigma1=3.0, sigma2=1.5, c1=(-5.5,-5.5), c2=(3.0,3.0), q_j=sigma_j^2
    Hand-chosen in Eq. (15) to create a narrow and a wide separated well.
  • Predator-prey interaction parameters A, l, c, C, sigma in V(d) = A=0.3, l=1.0, c=10^3, C=10.0, sigma=10.0
    Hand-tuned in Eq. (18) to satisfy heuristic conditions that produce push-out from narrow wells and oscillations in wide wells.
  • alpha_y (predator speed)
    Parameter in W in Eq. (17); must satisfy A > alpha_y; actual value used in simulation not stated.
assumptions (5)
  • domain assumption The Eyring formula is valid for the high-dimensional loss landscapes encountered in machine learning, i.e., landscapes with clear minima and saddles.
    Invoked in Section 2.2; the paper explicitly notes it may fail for landscapes without clear minima and saddles.
  • domain assumption Wide (flat) minima of the empirical risk correspond to solutions with better generalization (algorithmic stability).
    From references [37-40]; foundational to interpreting wide-minimum selection as overfitting control.
  • domain assumption The second contribution to the GAN functional V behaves approximately as the negative inverse KL divergence between discriminator and generator.
    Stated in Section 3.1 as qualitative behavior from [25]; used to argue that small KL near narrow maxima drives the predator-prey interaction.
  • domain assumption GAN training dynamics can be modeled by the SDE system (9)-(10) with independent Gaussian noise on discriminator and generator parameters.
    Assumed in Section 3.1; no derivation from the actual stochastic gradient updates used in GAN training.
  • ad hoc to paper The predator-prey interaction potential (18), with its piecewise distance-dependent behavior, is an appropriate stand-in for the GAN minimax interaction.
    Section 5.3 constructs V(d) with conditions chosen precisely to push out of narrow wells and oscillate in wide wells; this is a modeling choice specific to this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Control of Overfitting with Physics." pith.science (2026). https://pith.science/paper/MPZKC47P

@misc{pith2026241210716,
  author       = {Pith},
  title        = {Pith review of: Control of Overfitting with Physics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MPZKC47P}},
  note         = {Machine review of arXiv:2412.10716}
}
read the original abstract

While there are many works on the applications of machine learning, not so many of them are trying to understand the theoretical justifications to explain their efficiency. In this work, overfitting control (or generalization property) in machine learning is explained using analogies from physics and biology. For stochastic gradient Langevin dynamics, we show that the Eyring formula of kinetic theory allows to control overfitting in the algorithmic stability approach - when wide minima of the risk function with low free energy correspond to low overfitting. For the generative adversarial network (GAN) model, we establish an analogy between GAN and the predator-prey model in biology. An application of this analogy allows us to explain the selection of wide likelihood maxima and overfitting reduction for GANs.

Figures

Figures reproduced from arXiv: 2412.10716 by the authors.

Figure 1
Figure 1. Thermal plot of the function L (left) and its gradient field (right). Red dot in the center of the gradient field plot shows the starting point (0, 0). We consider 15 values of temperature T uniformly spanned on [0, 0.8]. For each value of T, we run the SGLD starting from the same initial point x0 with a maximal number of iterations K = 2000 and select those runs which in no more than K iterations converge to one or… view at source ↗
Figure 2
Figure 2. Fraction of the runs of the SGLD starting at the point [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Fraction of the points which converge to extrema vs. iteration number plotted for several [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Absolute value of the norm of the vector function [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: The limiting unstable oscillations around the extremum point for the predator–prey [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Simulation of the GAN process for the potentials [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Predator-Prey Model: Driven Hunt for Accelerated Grokking

    cs.NE 2025-09 conditional novelty 4.0 of 10

    A predator-prey two-agent optimizer accelerates the post-memorization grokking phase by tens to a hundred times in gradient calls on modular arithmetic and MNIST, but still requires standard pre-training to memorization.

Reference graph

Works this paper leans on

44 extracted references · 33 canonical work pages · cited by 1 Pith paper

  1. [1]

    Computing machinery and intelligence

    Turing, A. Computing machinery and intelligence. Mind 1950, 59, 433–460

    work page 1950

  2. [2]

    The Computer and the Brain, 1st ed.; Yale University Press: New Haven, USA, 1958

    von Neumann, J. The Computer and the Brain, 1st ed.; Yale University Press: New Haven, USA, 1958

    work page 1958

  3. [3]

    Complexity vs

    Manin, Y.I. Complexity vs. energy: Theory of computation and theoretical physics. J. Phys. Conf. Ser. 2014, 532, 012018. https://doi.org/10.1088/1742-6596/532/1/012018

    work page doi:10.1088/1742-6596/532/1/012018 2014

  4. [4]

    Genetic Algorithms in Search, Optimization and Machine Learning, 1st ed.; Addison-Wesley: Boston, MA, USA, 1989

    Goldberg, D.E. Genetic Algorithms in Search, Optimization and Machine Learning, 1st ed.; Addison-Wesley: Boston, MA, USA, 1989

    work page 1989

  5. [5]

    Evolutionary algorithms: A critical review and its future prospects

    Vikhar, P.A. Evolutionary algorithms: A critical review and its future prospects. In Proceedings of the International Conference on Global Trends in Signal Processing, Information Computing and Communication, Jalgaon, India, 22–24 December 2016. https://doi.org/10.1109/ICGTSPICC.2016.7955308

    work page arXiv 2016

  6. [6]

    Neural networks and physical systems with emergent collective computational abilities

    Hopfield, J.J. Neural networks and physical systems with emergent collective computational abilities. Proc. Natl. Acad. Sci. USA1982, 79, 2554–2558, https://doi.org/10.1073/ pnas.79.8.2554. 17

    work page

  7. [7]

    A learning algorithm for Boltzmann machines

    Ackley, D.H.; Hinton, G.E.; Sejnowski, T.J. A learning algorithm for Boltzmann machines. Cogn. Sci.1985, 9, 147–169

    work page 1985

  8. [8]

    Lotka-Volterra Model with Mutations and Generative Adversarial Networks

    Kozyrev, S.V. Lotka–Volterra Model with Mutations and Generative Adversarial Networks. Theor. Math. Phys.2024, 218, 276–284, https://arxiv.org/abs/2402.19035

    work page Pith review arXiv 2024

Show all 44 references
  1. [9]

    Transformers as a Physical Model in AI

    Kozyrev, S.V. Transformers as a Physical Model in AI. Lobachevskii J. Math. 2024, 45, 710–717. https://doi.org/10.1134/S1995080224600353

    2024 doi

  2. [10]

    A Tutorial on Energy-Based Learning; MIT Press: Cambridge, MA, USA, 2006

    LeCun, Y.; Chopra, S.; Hadsell, R.; Ranzato, M.A.; Huang, F.J. A Tutorial on Energy-Based Learning; MIT Press: Cambridge, MA, USA, 2006

    2006

  3. [11]

    Borgs, C.; Chayes, J.; Sagun, L.; Zecchina, R

    Chaudhari, P.; Choromanska, A.; Soatto, S.; LeCun, Y.; Baldassi, C. ; Borgs, C.; Chayes, J.; Sagun, L.; Zecchina, R. Entropy-SGD: Biasing Gradient Descent Into Wide Valleys. In Proceedings of the ICLR, Toulon, France, 24–26 April 2017. https://arxiv.org/abs/ arXiv:1611.01838

    2017 arXiv

  4. [12]

    Towards physical principles of biological evolution

    Katsnelson, M.I.; Wolf, Y.I; Koonin, E.V. Towards physical principles of biological evolution. Phys. Scr. 2018, 93 043001. https://arxiv.org/abs/1709.00284

    2018 arXiv

  5. [13]

    Towards a theory of evolution as multilevel learning

    Vanchurin, V.; Wolf, Yu.I.; Katsnelson, M.I.; Koonin, E.V. Towards a theory of evolution as multilevel learning. Proc. Natl. Acad. Sci. USA2022, 119, e2120037119. https://doi. org/10.1073/pnas.2120037119

    doi

  6. [15]

    Physics-informed neural net- works for quantum control

    Norambuena, A.; Mattheakis, M.; Gonz´ alez, F.J.; Coto, R. Physics-informed neural net- works for quantum control. Phys. Rev. Lett. 2024, 132, 010801. https://doi.org/10. 48550/ARXIV.2206.06287

    2024 doi arXiv

  7. [16]

    Incoherent Control of Quantum Systems With Wavefunction-Controllable Subspaces via Quantum Reinforcement Learning

    Dong, D.; Chen, C.; Tarn, T.-J.; Pechen, A.; Rabitz, H. Incoherent Control of Quantum Systems With Wavefunction-Controllable Subspaces via Quantum Reinforcement Learning. IEEE Trans. Syst. Man. Cybern. B Cybern.2008, 38, 957–962, https://doi.org/10.1109/ TSMCB.2008.926603

    2008

  8. [17]

    Quantum Reinforcement Learning

    Dong, D.; Chen, C.; Li, H.; Tarn, T.-J. Quantum Reinforcement Learning. IEEE Trans. Syst. Man. Cybern. B Cybern.2008, 38, 1207–1220. https://doi.org/10.1109/TSMCB. 2008.925743

    2008

  9. [18]

    Teaching the environment to control quantum systems

    Pechen, A.; Rabitz, H. Teaching the environment to control quantum systems. Phys. Rev. A 2006, 73, 062102. https://doi.org/10.1103/PhysRevA.73.062102

    2006 doi

  10. [19]

    Quan- tum Machine Learning

    Biamonte, J.; Wittek, P.; Pancotti, N.; Rebentrost, P.; Wiebe, N.; Lloyd, S. Quan- tum Machine Learning. Nature 2017, 549, 7671, 195–202. https://doi.org/10.1038/ nature23474. 18

    2017

  11. [20]

    M.; Buchhold, M.; Diehl, S

    Sieberer, L. M.; Buchhold, M.; Diehl, S. Keldysh field theory for driven open quantum systems. Rep. Prog. Phys.2016, 79, 096001

    2016

  12. [21]

    Complex quantum networks: A topical review

    Nokkala, J.; Piilo, J.;Bianconi, G. Complex quantum networks: A topical review. J. Phys. A Math. Theor.2024, 57, 233001

    2024

  13. [22]

    Quantum Brownian motion model for the stock market

    Meng, X.; Zhang, J.-W.; Guo, H. Quantum Brownian motion model for the stock market. Physica A2016, 452, 281–288

  14. [23]

    Delegated quantum neural networks for encrypted data

    Sun, W.; Chang, Y.; Wang, D.; Zhang, S.; Yan, L. Delegated quantum neural networks for encrypted data. Phys. Scr. 2024, 99, 05510. https://doi.org/10.1088/1402-4896/ ad348f

    2024 doi

  15. [24]

    The Activated Complex in Chemical Reactions

    Eyring, H. The Activated Complex in Chemical Reactions. J. Chem. Phys.1935, 3, 107–115. https://doi.org/10.1063/1.1749604

    1935 doi

  16. [25]

    Generative Adversarial Nets

    Goodfellow, I.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; Bengio, Y. Generative Adversarial Nets. In Proceedings of the NIPS, Montr´ eal, QC, Canada, 8–13 December 2014. https://arxiv.org/abs/arXiv:1406.2661

    2014 arXiv

  17. [26]

    Control of the von Neumann Entropy for an Open Two-Qubit System Using Coherent and Incoherent Drives

    Morzhin, O.V.; Pechen, A.N. Control of the von Neumann Entropy for an Open Two-Qubit System Using Coherent and Incoherent Drives. Entropy 2023, 26, 36. https://doi.org/ 10.3390/e26010036

    2023 doi

  18. [27]

    Training Generative Adversarial Networks by Solving Ordinary Differential Equations

    Qin, C.; Wu, Y.; Springenberg, J.T.; Brock, A.; Donahue, J.; Lillicrap, T.P.; Kohli, P. Training Generative Adversarial Networks by Solving Ordinary Differential Equations. In Proceedings of the NeurIPS, Vancouver, BC, Canada, 6–12 December 2020.https://arxiv. org/abs/arXiv:2010.15040

    2020 arXiv

  19. [28]

    Functional Space Analysis of Local GAN Conver- gence

    Khrulkov, V.; Babenko, A.; Oseledets, I. Functional Space Analysis of Local GAN Conver- gence. In Proceedings of the 38-th International Conference on Machine Learning, Virtual, 18–24 July 2021. https://arxiv.org/abs/arXiv:2102.04448

    2021 arXiv

  20. [29]

    Gradient descent GAN optimization is locally stable

    Nagarajan, V.; Kolter, J.Z. Gradient descent GAN optimization is locally stable. In Pro- ceedings of the 31st Conference on Neural Information Processing Systems, Long Beach, CA, USA, 4–9 December 2017. https://arxiv.org/abs/arXiv:1706.04156

    2017 arXiv

  21. [30]

    Correlation functions and computer simulations

    Parisi, G. Correlation functions and computer simulations. Nucl. Phys. B1981, 180, 378–

  22. [31]

    Correlation functions and computer simulations II

    Parisi, G. Correlation functions and computer simulations II. Nucl. Phys. B 1982, 205, 337–344. https://doi.org/10.1016/0550-3213(82)90360-1

    1982 doi

  23. [32]

    Diffusions for Global Optimization

    Geman, S.; Hwang, C.-R. Diffusions for Global Optimization. SIAM J. Control Optim.1986, 24, 1031–1043. https://doi.org/10.1137/0324060

    1986 doi

  24. [33]

    Bayesian Learning via Stochastic Gradient Langevin Dynamics

    Welling, M.; Teh, Y.W. Bayesian Learning via Stochastic Gradient Langevin Dynamics. In Proceedings of the 28-th International Conference on Machine Learning, Bellevue, DC, USA, 28 June–2 July 2011. 19

    2011

  25. [34]

    The Arrhenius formula in kinetic theory and Witten’s spectral asymptotics

    Kozyrev, S.V.; Volovich, I.V. The Arrhenius formula in kinetic theory and Witten’s spectral asymptotics. J. Phys. A2011, 44, 215202. https://arxiv.org/abs/1008.4487

    arXiv

  26. [35]

    High-Resolution Image Syn- thesis with Latent Diffusion Models

    Rombach, R.; Blattmann, A.; Lorenz D.; Esser P.; Ommer B. High-Resolution Image Syn- thesis with Latent Diffusion Models. arXiv 2022, https://arxiv.org/abs/2112.10752

    2022 arXiv

  27. [36]

    On the Design Fundamentals of Diffusion Models: A Survey

    Chang, Z.; Koulieris, G.A.; Shum H.P.H. On the Design Fundamentals of Diffusion Models: A Survey. arXiv 2023, https://arxiv.org/abs/2306.04542

    2023 arXiv

  28. [37]

    Stability and Generalization

    Bousquet, O.; Elisseeff, A. Stability and Generalization. J. Mach. Learn. Res. 2002, 2, 499–526

    2002

  29. [38]

    Almost-everywhere algorithmic stability and generalization error

    Kutin, S.; Niyogi, P. Almost-everywhere algorithmic stability and generalization error. In Proceedings of the 18-th Conference on Uncertainty in Artificial Intelligence, Edmonton, AB, Canada, 1–4 August 2002. https://arxiv.org/abs/arXiv:1301.0579

    2002 arXiv

  30. [39]

    General conditions for predictivity in learning theory

    Poggio, T.; Rifkin, R.; Mukherjee, S.; Niyogi, P. General conditions for predictivity in learning theory. Nature 2004, 428, 419–422. https://doi.org/10.1038/nature02341

    2004 doi

  31. [40]

    Flat Minima

    Hochreiter, S.; Schmidhuber, J. Flat Minima. Neural Comput.1997, 9, 1–42. https://doi. org/10.1162/neco.1997.9.1.1

    1997 doi

  32. [41]

    Geometric Characterization of the Eyring– Kramers Formula

    Avelin, B.; Julin, V.; Viitasaari, L. Geometric Characterization of the Eyring– Kramers Formula. Commun. Math. Phys.2023, 404, 401–437. https://doi.org/10.1007/ s00220-023-04845-z

    2023

  33. [42]

    Branching Processes, 1st ed.; Nauka: Moscow, Russia, 1971

    Sevastyanov, B.A. Branching Processes, 1st ed.; Nauka: Moscow, Russia, 1971

    1971

  34. [43]

    Branching Processes: Variation, Growth, and Extinc- tion of Populations; Cambridge University Press: Cambridge, UK, 2005

    Haccou, P.; Jagers, P.; Vatutin, V.A. Branching Processes: Variation, Growth, and Extinc- tion of Populations; Cambridge University Press: Cambridge, UK, 2005

    2005

  35. [44]

    A Course in the Theory of Stochastic Processes; McGraw-Hill International: New York, NY, USA, 1981

    Wentzell, A.D. A Course in the Theory of Stochastic Processes; McGraw-Hill International: New York, NY, USA, 1981. 20

    1981

  36. [384]

    https://doi.org/10.1016/0550-3213(81)90056-0

    doi

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.