Pith. sign in

REVIEW 3 major objections 5 minor 42 references

Resonant Machine Learning Based on Complex Growth Transform Dynamical Systems

T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Pith's one-line read This paper proposes a complex-domain growth-transform network that reaches an optimum with every node's voltage and current at ±90°, storing trained parameters as self-sustained oscillations with zero active power.

desk verdict Interesting extension of growth transforms to complex phasors with an active/reactive power interpretation, but the central convergence proof is unfinished and the physical claim is partly built into the objective. read the letter →

arxiv 1908.05377 v3 pith:WH5BX43X submitted 2019-08-15 cs.LG cs.SYeess.SYstat.ML

classification cs.LGcs.SYeess.SYstat.ML MSC 68T0568T0734C15
keywords resonantmachinelearningcomplexgrowthtransformTellegen'stheoremactivepowerdissipationreactiveconservationone-classsupportvectorLCoscillatornetworkzerosteadystate
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 show that a machine learning model can be built like a resonant electrical circuit, so that after training it holds its solution in self-sustained oscillations instead of dissipating energy. The authors reformulate learning as an optimization over complex voltage and current phasors, using Tellegen's theorem to separate active power (dissipated) from reactive power (stored), and they require the total reactive power to stay zero. They then introduce a complex-domain growth-transform dynamical system and prove, in Theorem 1, that it converges to the optimal point of the regularized objective with zero active power, meaning every nonzero node settles at a 90-degree voltage-current phase. If correct, this gives a physical route to energy-efficient analog machine learning, illustrated here with one-class support vector machines whose support vectors become resonant LC tanks.

What carries the argument

The central object is the complex growth transform dynamical system, in which each voltage evolves as $\partial V_i/\partial t = j\omega\sigma_{V_i}V_i - \Delta\sigma_{V_i}V_i$, each current evolves similarly with an added phase frequency, and the phase $\phi_i$ obeys a first-order filter $\tau_i\omega_{\phi_i} + \phi_i = g_{\phi_i}$. The load-bearing identity is Tellegen's theorem, which gives total complex power zero and lets the authors replace the zero-reactive-power constraint with conservation of reactive energy $\sum_i (\frac12 C_i|V_i|^2 + \frac12 L_i|I_i|^2) = E_0$. The phasor magnitudes carry the learning objective $H$, while the phase regularizer $\beta\sum_i |V_i|^2|I_i|^2\cos^2\phi_i$ shapes the trajectory so that steady state is reached at resonance, with $\phi_i = \pm\pi/2$ at every active node.

What would settle it

Run the update equations (15)-(18) on a two-node network from many random initial phases and record whether $\sum_i |V_i||I_i|\cos\phi_i$ converges to zero in every trial; a trajectory ending at any phase other than $\pm\pi/2$ with nonzero $|V_i||I_i|$ would disprove Theorem 1.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that the system of nonlinear dynamical equations (15)-(18) converges to the optimal point of problem (14) in steady state. The magnitude updates follow a growth transform that monotonically decreases a Lipschitz objective, and the phase updates drive each phase angle to $\pm\pi/2$ whenever the node carries nonzero voltage and current. In steady state the active power, $\sum_i |V_i||I_i|\cos\phi_i$, is zero, so the network stores learned parameters as limit-cycle oscillations sustained by capacitive and inductive reactive energy. Applied to a one-class SVM with $\alpha_i = |V_i|^2 + |I_i|^2$, support vectors correspond to LC tanks at resonance and interior points to grounded sinks, and the resonant model is reported to dissipate less active power than the non-resonant variant.

Load-bearing premise

The zero-power conclusion rests on the assumption that the rule for adjusting the voltage-current phase always lowers the cost and always settles at a 90-degree offset; if some other phase offset can be stable, the trained network will keep dissipating power.

Editorial extensions

If this is right

  • A trained one-class SVM can in principle be stored as self-sustained LC oscillations, drawing no active power from the supply in steady state.
  • The annealing schedule for $\beta$ trades convergence speed against active-power dissipation during learning, allowing different operating points between fast learning and low energy use.
  • The same construction applies to any learning problem of the form (26) over a probability simplex, including probabilistic models, and to coupled-oscillator networks for optimization and pattern matching.
  • The network remains at resonance under a global phase shift or a relative phase shift between voltage and current, so the learned solution is invariant to common phase drifts.
  • The second-order dynamics split into a conservative limit-cycle term and a dissipative term, which the paper reads as a stable limit cycle robust to small perturbations.

Reading between the lines

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

  • If the zero-active-power steady state is robust in hardware, energy-based analog inference could run continuously on stored reactive energy, making always-on edge classifiers plausible.
  • The phase variable offers a physical substrate for synchronization; a testable extension is whether the resonant SVM's support vectors phase-lock, linking classification margins to oscillator synchronization.
  • A direct numerical check of the phase subsystem alone, starting from many random initial phases and recording where $\phi_i$ settles, would isolate the unproven Condition 3 without building any circuit.
  • Because the learned solution lives in the frequency and phase of oscillations, perturbation recovery on the limit cycle suggests a form of analog error correction absent in dissipative energy-based models.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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. This manuscript proposes a complex-domain, continuous-time growth-transform dynamical system for solving constrained optimization problems whose variables are interpreted as voltage and current phasors of an electrical network. The central idea is to augment a magnitude-only objective H(|V_i|,|I_i|) with an active-power dissipation regularizer β∑|V_i|^2|I_i|^2 cos^2 φ_i and a conserved reactive-energy constraint ∑(|V_i|^2+|I_i|^2)=1. The paper states Theorem 1 claiming that the coupled magnitude-phase dynamics (15)-(18) converge to the optimum of Equation (14) and reach a steady state with zero active power, i.e., φ_i=±π/2 at every node with nonzero |V_i||I_i|. It then applies the framework to one-class SVM on synthetic and real datasets, reporting dissipation profiles and support-vector counts.

Significance. The potential significance is real if the central convergence theorem can be established. The framework offers a conceptually novel energy-based learning formulation in which learned parameters are encoded in sustained LC oscillations, active power is dissipated only transiently, and an annealing schedule trades off convergence speed against dissipation. The paper also provides a general mapping from probabilistic optimization problems to phasor magnitudes, and its simulation study covers multiple annealing schedules, random initializations, and several benchmark datasets. The writing is generally clear. However, the main theorem currently rests on an unproved Lyapunov inequality for the coupled phase-magnitude dynamics, and the physical equivalence used to motivate the conservation constraint is stated too strongly. These issues are central rather than cosmetic.

major comments (3)
  1. [Appendix C, Condition 3 (Eqs. (15)-(18), (57)-(59))] The proof of the phase update is asserted, not derived. The text says "We can then apply the growth transform dynamical system [7] to get" and then jumps to Eq. (57) and the resulting Eq. (17); no Lyapunov inequality is established for the coupled system in which the magnitude updates (15)-(16) depend on φ through ∂L/∂V_i and ∂L/∂I_i, and the phase update (17) depends on the magnitudes through ∂L/∂φ_i. Since the steady-state zero-active-power claim is exactly the statement that φ_i=±π/2 for every node with |V_i||I_i|≠0, this missing monotonicity argument is load-bearing for Theorem 1. Please supply the missing derivation or explicitly weaken the theorem to a conjecture supported by simulations.
  2. [Section III, Eqs. (5)-(6)] The claim that zero total reactive power, ∑ [V_i(C_i dV_i/dt)* + (L_i dI_i/dt) I_i*]=0, is equivalent to conservation of reactive energy ∑(1/2 C_i|V_i|^2 + 1/2 L_i|I_i|^2)=E0 is not correct in general. In sinusoidal steady state the left side is an imaginary-power balance involving ω∑(L_i|I_i|^2 - C_i|V_i|^2), not the time derivative of the stored energy; the derivative of the stored energy is related to real power. This conflation affects the physical interpretation of the normalization constraint (9) and of the claimed electrical-resonance equivalence. The optimization can still be posed with (9) as a normalization constraint, but the stated equivalence should be corrected or qualified.
  3. [Appendix C, Lemma leading to Eqs. (50)-(53)] The continuous-time limit is not derived with the required rate normalization. The discrete update (52) contains the term V_{i,n-1} Δσ_{V_i,n-1}; passing to Δt→0 yields ∂V_i/∂t = jωσ_{V_i} V_i - Δσ_{V_i} V_i only if Δσ is O(Δt) (or an explicit time constant is introduced and the terms are scaled accordingly). No such normalization is specified, so the continuous-time dynamics (15)-(16) are only heuristically justified. Please provide a rigorous limiting argument or state the heuristic status of this step.
minor comments (5)
  1. [Table numbering, Section V] Table 2 is introduced for the growth-transform system in Section IV, and then a second Table 2 titled "Performance on real benchmark datasets" appears in Section V; the second table should be renumbered and all references updated.
  2. [Theorem 1 statement] The sum in the theorem statement, ∑_{n=1}^N |V_i||I_i| cos φ_i, uses index n in the summation but i in the summand; it should be ∑_{i=1}^N |V_i||I_i| cos φ_i.
  3. [Section III, Eq. (4) and Eq. (7)] Equation (4) defines D with |Re{V_i I_i^*}|^2 while later the same symbol is used for |V_i|^2|I_i|^2 cos^2 φ_i. Since φ_i is defined as the phase angle between V_i and I_i, these expressions agree, but the equivalence should be stated explicitly at the first occurrence.
  4. [Section V, Table 2] The real-data results report only counts of inliers, outliers, and support vectors, with no baseline accuracy, ROC, or AUC comparison against a standard one-class SVM; adding such a baseline would let the reader judge whether the resonant regularizer preserves decision quality.
  5. [Figures 6 and 8] The insets showing transient dynamics are very small; enlarging them would make the claimed convergence behavior easier to assess.

Circularity Check

2 steps flagged · score 5.0 of 10

Partial circularity: the zero-active-power steady state is baked into the objective via the cos²φ regularizer, and the phase-convergence proof leans on an asserted application of the authors' prior growth-transform result.

  1. self definitional [Section III, after Eq. (10), 'two important properties' bullet; Theorem 1]
    "If β is slowly annealed to a sufficiently high value, then φi→π/2 under steady state for i :|Vi||Ii|⁄= 0. This implies that network achieves zero active power dissipation in steady state."

    The regularizer in Eq. (10)/(14) is βΣ|V_i|^2|I_i|^2 cos²φ_i. Its global minima occur exactly where cos²φ_i = 0, i.e., φ_i = ±π/2, and at those points Σ|V_i||I_i|cosφ_i = 0 by elementary trigonometry. Thus the theorem's zero-active-power steady state is a property designed into the optimization objective, not an emergent consequence of the dynamics or of Tellegen/resonance physics. The dynamical system is claimed to converge to a minimizer of this pre-designed objective, so the zero-dissipation conclusion is an input of the problem formulation rather than an independent prediction.

  2. self citation load bearing [Appendix C, Condition 3, phase update derivation]
    "We can split φi as φi =φ+i−φ−i, φ+i,φ−i>0, which implies that φ+i+φ−i=π. We can then apply the growth transform dynamical system [7] to get ... This implies that the voltage and current phasors corresponding to the i−th node in the network may be phase shifted by an additional amount φi with respect to the absolute reference."

    The convergence of the phase variables is the load-bearing step for the zero-dissipation claim, and it is justified by applying the authors' own growth-transform system [7] to the decomposed phase pair. The text then asserts 'This implies' the desired steady state without deriving a Lyapunov inequality for the actual coupled system (15)-(18), where magnitude updates depend on φ and the phase update depends on magnitudes. The conclusion φ_i = ±π/2 is exactly what the phase regularizer was constructed to enforce, so the self-citation [7] is invoked to close the gap between the constructed objective and the theorem's headline physical result.

full rationale

The paper's core mathematical proposal—the complex growth-transform dynamical system (15)-(18)—is not circular in the statistical sense: no parameter is fitted to data and then renamed a prediction, and the SVM experiments are simulations of the proposed dynamics on standard benchmarks. However, the headline outcome, steady-state zero active-power dissipation, is placed into the objective by construction: Eq. (10)/(14) includes βΣ|V_i|^2|I_i|^2cos²φ_i, whose minimizers are exactly the quadrature points where Σ|V_i||I_i|cosφ_i = 0. The dynamics only need to reach that pre-designed optimum. Furthermore, the proof that the coupled system decreases L is incomplete: Appendix C, Condition 3 applies the authors' earlier growth-transform result [7] to the phase pair and then asserts the desired convergence for the full coupled magnitude-phase system; no monotonicity proof for the coupled system is supplied. This is partly a proof gap and partly a self-citation load-bearing step, but combined with the by-construction regularizer it makes the zero-dissipation claim substantially less independent than the paper's framing suggests. Overall, partial circularity with genuine independent content in the dynamical-system derivation, so a moderate score rather than a full reduction.

Assumptions & free parameters 7 free parameters · 6 assumptions · 2 invented entities

The framework rests on standard circuit theorems and the authors' prior growth transform results, but adds a nonstandard equivalence (Eqs. 5-6) and a phase regularizer that directly encodes the zero-dissipation target. Several hyperparameters and dynamical constants are chosen by hand and not reported.

free parameters (7)
  • beta (active-power regularization weight) = 1 or 10 in experiments; logistic and switching annealing schedules
    Controls the trade-off between the original objective H and the phase-dependent dissipation penalty D; central to the resonance claim.
  • angular frequency omega = pi/10, pi/4, pi/20, pi/8 in experiments
    Common oscillation frequency of all phasors in the dynamical system; chosen by hand.
  • kernel width sigma = 1, 10, 20 for synthetic datasets
    Gaussian kernel parameter for the one-class SVM; chosen per dataset.
  • nu (OC-SVM regularization) = 0.1
    Controls the fraction of outliers; used in all experiments.
  • lambda (growth-transform positivity constant) = not reported
    Must be large enough to keep update terms positive (Table 2, Appendix C); no numerical values are given.
  • time constants tau_i for phase updates = not reported
    Determine phase evolution speed in Eq. (17); values are not stated.
  • annealing hyperparameters k and t0 = not reported numerically
    Logistic beta schedule steepness and midpoint; chosen without stated values.
assumptions (6)
  • standard math Tellegen's theorem: total complex power over all network nodes is zero (Eq. 2).
    Basis for the relation PT+jQT = -PN-jQN and for the power decomposition used throughout Section III.
  • standard math Baum-Eagon inequality / growth transform convergence for Lipschitz continuous objective functions on the probability simplex.
    Core machinery for the magnitude update equations, cited to [7], [40], [41]; the present paper extends it to complex variables.
  • standard math Wirtinger calculus for derivatives of real-valued functions of complex variables.
    Used in Appendix C to express partial derivatives with respect to V_i and I_i.
  • ad hoc to paper Zero reactive power is equivalent to conservation of total reactive energy (Eqs. 5-6).
    The paper treats these as equivalent, but in sinusoidal steady state Q=0 only equates average stored electric and magnetic energies; total energy constancy is an additional postulate that produces the normalization constraint.
  • domain assumption The learning objective H depends only on magnitudes |V_i| and |I_i|, allowing phase angles to be controlled independently.
    Decouples magnitude optimization from phase dynamics, as in Eq. (10); valid if H is constructed this way.
  • domain assumption Every learning parameter alpha_i can be encoded as |V_i|^2 + |I_i|^2 with V_i and I_i independent (Eq. 27).
    Maps the original SVM optimization to the complex LC network; the dynamics exploit the redundancy in this representation.
invented entities (2)
  • Reactive energy E0 (or its normalized version) as a conserved quantity
    purpose: Defines the feasible manifold and justifies the normalization sum(|V|^2+|I|^2)=1; supports the claim of self-sustained storage.
    No physical measurement or independent verification; introduced as a modeling postulate.
  • Resonant LC tank interpretation of support vectors
    purpose: Represents each support vector as a node with voltage and current in phase quadrature, storing learned parameters as sustained oscillations.
    Interpretation of the phase-regularized solution; no circuit is built or measured.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Resonant Machine Learning Based on Complex Growth Transform Dynamical Systems." pith.science (2026). https://pith.science/paper/WH5BX43X

@misc{pith2026190805377,
  author       = {Pith},
  title        = {Pith review of: Resonant Machine Learning Based on Complex Growth Transform Dynamical Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WH5BX43X}},
  note         = {Machine review of arXiv:1908.05377}
}
read the original abstract

Traditional energy-based learning models associate a single energy metric to each configuration of variables involved in the underlying optimization process. Such models associate the lowest energy state to the optimal configuration of variables under consideration, and are thus inherently dissipative. In this paper we propose an energy-efficient learning framework that exploits structural and functional similarities between a machine learning network and a general electrical network satisfying the Tellegen's theorem. In contrast to the standard energy-based models, the proposed formulation associates two energy components, namely, active and reactive energy to the network. This ensures that the network's active-power is dissipated only during the process of learning, whereas the reactive-power is maintained to be zero at all times. As a result, in steady-state, the learned parameters are stored and self-sustained by electrical resonance determined by the network's nodal inductances and capacitances. Based on this approach, this paper introduces three novel concepts: (a) A learning framework where the network's active-power dissipation is used as a regularization for a learning objective function that is subjected to zero total reactive-power constraint; (b) A dynamical system based on complex-domain, continuous-time growth transforms which optimizes the learning objective function and drives the network towards electrical resonance under steady-state operation; and (c) An annealing procedure that controls the trade-off between active-power dissipation and the speed of convergence. As a representative example, we show how the proposed framework can be used for designing resonant support vector machines (SVMs), where we show that the support-vectors correspond to an LC network with self-sustained oscillations.

Figures

Figures reproduced from arXiv: 1908.05377 by the authors.

Figure 1
Figure 1. (a) The total or apparent power SN of the electrical network comprising of the active-power PN or the dissipated power and the reactive￾power QN or the power used for energy-storage. The goal of the proposed resonant learning framework: (b) minimizing active-power PN during learning and ensuring PN = 0, post-learning or steady-state; and (c) maintaining QN = 0 in learning and post-learning phases. even if the soluti… view at source ↗
Figure 2
Figure 2. Equivalent network model comprising of N electrical nodes, with an inductance and capacitance element associated with each of the nodes. (a) Learning is equivalent to changing the values of inductive and capacitive elements; and (b) in steady-state the network is driven into electrical resonance. Conventional Complex 0 1 Initial Final 𝑉௜ ଶ = 𝑐 Im 𝑉௜ Re 𝑉௜ [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Illustration showing that operating in the complex domain allows [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Cost function L1 plotted for different values of the hyperparameter β: (a) β = 0; (b) β = 1 and (c) β = 10 [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Comparison of the resonant optimization model ( [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Comparison of the performance of the resonant model [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Comparison of the active and reactive power dissipated at each node for the non-resonant model [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Comparison of the performance of the resonant model [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Robustness to random initialization: Comparison of (a) the time [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Comparison of the dissipation profiles of the resonant model [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Circuit based and phasor based representations for a one-class [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: LC tank resonator Consider the parallel LC tank circuit shown in [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 37 canonical work pages

  1. [7]

    Decentralized global optimization based on a growth transform dynamical system model,

    O. Chatterjee and S. Chakrabartty, “Decentralized global optimization based on a growth transform dynamical system model,” IEEE Trans- actions on Neural Networks and Learning Systems , vol. 29, no. 12, pp. 6052–6061, 2018

  2. [1]

    A tutorial on energy-based learning,

    Y . LeCun, S. Chopra, R. Hadsell, M. Ranzato, and F. Huang, “A tutorial on energy-based learning,” Predicting Structured Data , vol. 1, no. 0, 2006

  3. [2]

    Neural computation of decisions in optimization problems,

    J. J. Hopfield and D. W. Tank, “Neural computation of decisions in optimization problems,” Biological Cybernetics, vol. 52, no. 3, pp. 141– 152, 1985

  4. [3]

    Edminister, Theory and problems of electric circuits

    J. Edminister, Theory and problems of electric circuits . Schaum, 1965

  5. [4]

    M. E. V . Valkenburg, Network analysis. Prentice-Hall, 1964

  6. [5]

    Reactive power compensation technologies: State-of-the-art review,

    J. Dixon, L. Moran, J. Rodriguez, and R. Domke, “Reactive power compensation technologies: State-of-the-art review,” Proceedings of the IEEE, vol. 93, no. 12, pp. 2144–2164, 2005

  7. [6]

    Hofmann, J

    W. Hofmann, J. Schlabbach, and W. Just, Reactive power compensation: a practical guide . John Wiley & Sons, 2012. 15

  8. [8]

    Support vector method for novelty detection,

    B. Sch ¨olkopf, R. C. Williamson, A. J. Smola, J. Shawe-Taylor, and J. C. Platt, “Support vector method for novelty detection,” in Advances in Neural Information Processing Systems , pp. 582–588, 2000

Show all 42 references
  1. [9]

    From kuramoto to crawford: exploring the onset of synchronization in populations of coupled oscillators,

    S. H. Strogatz, “From kuramoto to crawford: exploring the onset of synchronization in populations of coupled oscillators,” Physica D: Nonlinear Phenomena, vol. 143, no. 1-4, pp. 1–20, 2000

  2. [10]

    The character of physical law,

    R. Feynman, “The character of physical law,” 2017

  3. [11]

    A unified energy-based framework for unsupervised learning,

    Y .-L. Boureau, S. Chopra, Y . LeCun, et al. , “A unified energy-based framework for unsupervised learning,” in Artificial Intelligence and Statistics, pp. 371–379, 2007

  4. [12]

    Optimal perceptual inference,

    G. E. Hinton and T. J. Sejnowski, “Optimal perceptual inference,” in Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pp. 448–453, 1983

  5. [13]

    Restricted boltzmann machines for collaborative filtering,

    R. Salakhutdinov, A. Mnih, and G. Hinton, “Restricted boltzmann machines for collaborative filtering,” in Proceedings of the 24th inter- national conference on Machine learning , pp. 791–798, ACM, 2007

  6. [14]

    Equilibrium propagation: Bridging the gap between energy-based models and backpropagation,

    B. Scellier and Y . Bengio, “Equilibrium propagation: Bridging the gap between energy-based models and backpropagation,” Frontiers in Computational Neuroscience, vol. 11, p. 24, 2017

  7. [15]

    Learning deep energy models,

    J. Ngiam, Z. Chen, P. W. Koh, and A. Y . Ng, “Learning deep energy models,” in Proceedings of the 28 th International Conference on Machine Learning, 2011

  8. [16]

    Deep structured energy based models for anomaly detection,

    S. Zhai, Y . Cheng, W. Lu, and Z. Zhang, “Deep structured energy based models for anomaly detection,” in International Conference on Machine Learning, pp. 1100–1109, 2016

  9. [17]

    Reinforcement learning with deep energy-based policies,

    T. Haarnoja, H. Tang, P. Abbeel, and S. Levine, “Reinforcement learning with deep energy-based policies,” in Proceedings of the 34th Interna- tional Conference on Machine Learning-Volume 70 , pp. 1352–1361, JMLR. org, 2017

  10. [18]

    Energy-inspired models: Learning with sampler-induced distributions,

    J. Lawson, G. Tucker, B. Dai, and R. Ranganath, “Energy-inspired models: Learning with sampler-induced distributions,” in Advances in Neural Information Processing Systems , pp. 8499–8511, 2019

  11. [19]

    Hirose, Complex-valued neural networks: theories and applications , vol

    A. Hirose, Complex-valued neural networks: theories and applications , vol. 5. World Scientific, 2003

  12. [20]

    Deep complex networks,

    C. Trabelsi, O. Bilaniuk, Y . Zhang, D. Serdyuk, S. Subramanian, J. F. Santos, S. Mehri, N. Rostamzadeh, Y . Bengio, and C. J. Pal, “Deep complex networks,” arXiv preprint arXiv:1705.09792 , 2017

  13. [21]

    Deep quaternion networks,

    C. J. Gaudet and A. S. Maida, “Deep quaternion networks,” in 2018 International Joint Conference on Neural Networks (IJCNN) , pp. 1–8, IEEE, 2018

  14. [22]

    On complex valued convolutional neural networks,

    N. Guberman, “On complex valued convolutional neural networks,” arXiv preprint arXiv:1602.09046 , 2016

  15. [23]

    Complex support vector machines for regression and quaternary classification,

    P. Bouboulis, S. Theodoridis, C. Mavroforakis, and L. Evaggelatou- Dalla, “Complex support vector machines for regression and quaternary classification,” IEEE Transactions on Neural Networks and Learning Systems, vol. 26, no. 6, pp. 1260–1274, 2014

  16. [24]

    A mathematical motivation for complex-valued convolutional networks,

    J. Bruna, S. Chintala, Y . LeCun, S. Piantino, A. Szlam, and M. Tygert, “A mathematical motivation for complex-valued convolutional networks,” arXiv preprint arXiv:1503.03438 , 2015

  17. [25]

    Orthogonality of decision boundaries in complex-valued neural networks,

    T. Nitta, “Orthogonality of decision boundaries in complex-valued neural networks,” Neural Computation, vol. 16, no. 1, pp. 73–97, 2004

  18. [26]

    Complex-valued restricted boltzmann machine for direct learning of frequency spectra.,

    T. Nakashika, S. Takaki, and J. Yamagishi, “Complex-valued restricted boltzmann machine for direct learning of frequency spectra.,” in INTER- SPEECH, pp. 4021–4025, 2017

  19. [27]

    A general network theorem, with applications,

    B. D. Tellegen, “A general network theorem, with applications,” Philips Res Rep, vol. 7, pp. 256–269, 1952

  20. [28]

    The elements of statistical learning: data mining, inference and prediction,

    T. Hastie, R. Tibshirani, J. Friedman, and J. Franklin, “The elements of statistical learning: data mining, inference and prediction,” The Mathematical Intelligencer, vol. 27, no. 2, pp. 83–85, 2005

  21. [29]

    K. P. Murphy, Machine learning: a probabilistic perspective. MIT press, 2012

  22. [30]

    Probabilistic machine learning and artificial intelli- gence,

    Z. Ghahramani, “Probabilistic machine learning and artificial intelli- gence,” Nature, vol. 521, no. 7553, p. 452, 2015

  23. [31]

    Gini support vector machine: Quadratic entropy based robust multi-class probability regression,

    S. Chakrabartty and G. Cauwenberghs, “Gini support vector machine: Quadratic entropy based robust multi-class probability regression,” Jour- nal of Machine Learning Research , vol. 8, no. Apr, pp. 813–839, 2007

  24. [32]

    Support vector domain description,

    D. M. Tax and R. P. Duin, “Support vector domain description,” Pattern Recognition Letters, vol. 20, no. 11-13, pp. 1191–1199, 1999

  25. [33]

    Extended poly- nomial growth transforms for design and training of generalized support vector machines,

    A. Gangopadhyay, O. Chatterjee, and S. Chakrabartty, “Extended poly- nomial growth transforms for design and training of generalized support vector machines,” IEEE Transactions on Neural Networks and Learning Systems, vol. 29, no. 5, pp. 1961–1974, 2017

  26. [34]

    UCI machine learning repository

    D. Dua and C. Graff, “UCI machine learning repository.” http://archive. ics.uci.edu/ml, 2017

  27. [35]

    ODDS library

    S. Rayana, “ODDS library.” http://odds.cs.stonybrook.edu, 2016

  28. [36]

    Computing with networks of oscillatory dynamical systems,

    A. Raychowdhury, A. Parihar, G. H. Smith, V . Narayanan, G. Csaba, M. Jerry, W. Porod, and S. Datta, “Computing with networks of oscillatory dynamical systems,”Proceedings of the IEEE, vol. 107, no. 1, pp. 73–89, 2019

  29. [37]

    Neuromorphic computing with nanoscale spintronic oscillators,

    J. Torrejon, M. Riou, F. A. Araujo, S. Tsunegi, G. Khalsa, D. Querlioz, P. Bortolotti, V . Cros, K. Yakushiji, A. Fukushima,et al., “Neuromorphic computing with nanoscale spintronic oscillators,” Nature, vol. 547, no. 7664, p. 428, 2017

  30. [38]

    M. A. Nielsen and I. L. Chuang, Quantum computation and quantum information. Cambridge university press, 2010

  31. [39]

    Continuous- time optimization using sub-threshold current-mode growth transform circuits,

    A. Gangopadhyay, O. Chatterjee, and S. Chakrabartty, “Continuous- time optimization using sub-threshold current-mode growth transform circuits,” in 2018 IEEE 61st International Midwest Symposium on Circuits and Systems (MWSCAS) , pp. 246–249, IEEE, 2018

  32. [40]

    Growth transformations for functions on manifolds,

    L. E. Baum and G. Sell, “Growth transformations for functions on manifolds,” Pacific Journal of Mathematics , vol. 27, no. 2, pp. 211– 227, 1968

  33. [41]

    A generalization of the baum algorithm to rational objective functions,

    P. Gopalakrishnan, D. Kanevsky, A. Nadas, and D. Nahamoo, “A generalization of the baum algorithm to rational objective functions,” in International Conference on Acoustics, Speech, and Signal Processing , pp. 631–634, IEEE, 1989

  34. [42]

    Wirtinger calculus based gradient descent and levenberg-marquardt learning algorithms in complex-valued neural networks,

    M. F. Amin, M. I. Amin, A. Y . H. Al-Nuaimi, and K. Murase, “Wirtinger calculus based gradient descent and levenberg-marquardt learning algorithms in complex-valued neural networks,” inInternational Conference on Neural Information Processing , pp. 550–559, Springer, 2011

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.