Controllable Neural Architectures for Multi-Task Control

Fabio Pasqualetti; Giacomo Baggio; Sandro Zampieri; Umberto Casti

REVIEW 3 major objections 3 minor 31 references

Controllable Neural Architectures for Multi-Task Control

T0 review · 3 major / 3 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read A recurrent neural controller with fixed connectivity can replicate any single linear controller and approximate whole families of them by tuning only its bias vector.

desk verdict A genuinely new multi-task controller formulation with a formal gap: the H2 cost in problem (5) is only defined over stable linearizations, but stability is not enforced, leaving the optimization problem ill-posed as stated. read the letter →

arxiv 2501.19059 v1 pith:QORCLTIK submitted 2025-01-31 eess.SY cs.SY

Umberto Casti , Giacomo Baggio , Sandro Zampieri , Fabio Pasqualetti This is my paper · ORCID

classification eess.SYcs.SY

keywords multi-taskcontrolneuralcontrollerbiasadaptationlinearizeddynamicssmoothReLUH2normapproximationbalancedtruncationlowerbounds

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 argues that a single recurrent neural network can act as a controller for many different linear systems at once, provided the network's recurrent connectivity is fixed but its bias vector is allowed to vary from task to task. The bias changes the equilibrium point of the network, and the network's linearized dynamics around that equilibrium approximate the desired linear controller. The paper proves that any diagonal matrix $D$ with entries in $(0,1)^N$ can be realized by some bias vector, so with a free connectivity matrix $W$ the linearized dynamics $(-I+DW,B,C)$ can exactly reproduce the dynamics of any single linear controller. It then formulates multi-task approximation as minimizing the sum of $\mathcal{H}_2$ distances to $M$ desired controllers, gives analytic gradients for training, and proves upper and lower bounds on the achievable error. If correct, this gives a parameter-efficient way to store and switch among many controllers: one static network plus a low-dimensional bias per task.

What carries the argument

The load-bearing object is the bias-parametrized linearization of the neural controller. At an equilibrium $x_{\rm eq}=\Phi(Wx_{\rm eq}+d)$, the local dynamics are $\Sigma^L=(-I+DW,B,C)$ with $D=\operatorname{diag}(\phi_d(Wx_{\rm eq}+d))$ and $\phi_d(x)=1/(1+e^{-x})\in(0,1)$. Because $\phi_d$ is injective, Theorem 2.1 constructs $x_{\rm eq}$ and $d$ realizing any desired diagonal $D$, which converts the nonlinear controller into a linear-family approximation problem in the variables $(W,D_i,B,C)$. The upper bound is carried by balanced truncation: the $M$ systems are stacked into a parallel system, balanced and truncated, so that the shared-connectivity constraint appears as an output-averaging error $\Delta C$ plus a truncated-mode term; the lower bounds use time-zero impulse response differences and, for scalar systems, a Lambert-$W$ computation of the optimal pole placement.

What would settle it

Run a dense numerical search over $(W,D_1,D_2,B,C)$ with $N=1$ for two stable scalar systems with residues $r_1<r_2$ in problem (20); Theorem 4.3 asserts the optimum is exactly $\min\{A(1,2),-A(2,1),-r_1/a_1\}$, so any search point with lower cost would refute the lower bound, and for $M=2$ any failure to attain it would contradict the claimed equality.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes a representation and approximation result for a neural controller with smooth ReLU activation $\phi(x)=\ln(1+e^x)$. The central claim is that the linearized dynamics of the controller, evaluated at bias-dependent equilibria, form a flexible family: for any connectivity matrix $W$ and any diagonal $D\in(0,1)^N$ there exist an equilibrium $x_{\rm eq}$ and bias $d$ with $D=\operatorname{diag}(\phi_d(Wx_{\rm eq}+d))$, and therefore the linearized triple $\Sigma^L=(-I+DW,B,C)$ can match any desired linear system $(A,B,C)$ by taking $W=D^{-1}(A+I)$. For multiple tasks, the paper claims the minimization of $\sum_i \|\Sigma^D_i-\Sigma^L_i\|_2^2$ over shared $W,B,C$ and task-specific $D_i$ is tractable via explicit gradient formulas, and that the achievable error is governed by the number of modes the network can represent, the number of tasks, and how dissimilar the tasks are in a balanced realization. Theorem 4.1 states an upper bound in terms of the Hankel singular values of the parallel system and a term measuring output-matrix spread; Theorems 4.2 and 4.3 state lower bounds, with the scalar two-task case solved exactly.

Load-bearing premise

The error metric is the $\mathcal{H}_2$ norm, which is finite only when every linearized approximating system $(-I+D_iW,B,C)$ is stable; the optimization puts no stability constraint on those systems, and the linearization is only a local model around the bias-dependent equilibrium.

Editorial extensions

If this is right

A library of $M$ linear controllers can be replaced by one recurrent network plus $M$ bias vectors, reducing stored parameters when the network state dimension is below the controller dimension (Remark 1).
Any single desired linear controller can be embedded exactly as the linearization of the network at a suitable bias, so the architecture is a genuine superset of linear state-feedback controllers rather than only an approximation tool.
The multi-task error is controlled by the number of dominant modes of the stacked system and by the similarity of the desired output matrices in balanced coordinates; identical tasks incur only standard balanced-truncation error.
For two stable scalar systems with positive residues, the $\ell^1$ impulse-response error of the best scalar neural controller is known exactly, so in that setting the fundamental cost of sharing one controller is quantified.
The gradient formulas give a direct training procedure, and numerical experiments show the cost falling as the network dimension $N$ grows from 1 to 8 and rising as the number of tasks $M$ grows.

Reading between the lines

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

The paper treats each bias as selected from a pre-trained set; a natural extension, which the paper lists as future work, is to switch biases online during a trajectory, where the local linearization guarantee no longer covers transient behavior.
The parametrization theorem implies the achievable linearized dynamics are exactly those of the form $-I+DW$ with $D$ diagonal in $(0,1)^N$; this equivalence to diagonal precompensation suggests a direct link to structured sparse feedback design that the paper does not explore.
Because the $\mathcal{H}_2$ cost is only meaningful when every linearized system is stable, a practical training scheme would need a stability safeguard; the paper provides none, so its bounds apply to the subset of reachable stable linearizations.
One could test the parameter-efficiency claim empirically on embedded hardware by measuring update time and memory when switching between tasks, comparing the neural controller against storing separate LQR controllers.

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.

Desk Editor's Note

read the letter

Bottom line: this is a real contribution with a fixable gap. The core idea—one fixed connectivity matrix W with task-specific bias vectors d—is new relative to the cited single-model approximation work [22], and Theorem 2.1's universal realizability of any diagonal D is clean and correct. The gradient formulas in Theorem 3.1 are standard but correctly derived, and the balanced-truncation upper bound in Theorem 4.1 is a reasonable constructive argument. The lower bounds in Theorems 4.2 and 4.3 are more specialized, but they are honest attempts at fundamental limits.

The soft spots are real but not fatal. The main one is exactly what the stress-test note flagged: problem (5) minimizes a sum of H2 norms without restricting the approximating systems -I + D_i W to be stable. The H2 norm of the error system is finite only if those matrices are Hurwitz, so the domain of the optimization problem is not the set of all feasible variables, and the gradient formulas rely on Gramians that exist only for stable error systems. The paper even admits in Section III that the gradient procedure 'offers no stability or performance guarantees.' This is a formal ill-posedness in the problem statement, not just a numerical nuisance. It can be fixed by introducing stability constraints or an extended-valued cost, and the existence of stable approximants is partially rescued by the constructive proof of Theorem 4.1. But as written, the training problem is incomplete. Second, the proof of Theorem 4.3 explicitly omits two cases with a hand-wave; that is a minor completeness issue, but it needs to be filled in.

The numerical experiments are illustrative: no code or data are provided, and the plots are suggestive rather than conclusive. For a theory paper that is acceptable, though one reproducible example would strengthen it.

Overall, the central claim—that a fixed connectivity matrix with varying bias can approximate multiple linear controllers with explicit bounds—holds up in the theory sections. The well-posedness issue is genuine, but it is repairable and does not undermine the main conceptual contribution. This paper deserves a serious referee; I would send it out with expectations of major revisions, primarily to address the stability constraint in (5) and complete the proof of Theorem 4.3.

Referee Report

3 major / 3 minor

Summary. This paper proposes a nonlinear neural controller of the form ẋ = -x + Φ(Wx + d) + Bu, y = Cx, with a fixed connectivity matrix W and task-dependent bias vector d, and studies its ability to approximate a set of M desired linear controllers via the linearized dynamics around bias-dependent equilibria. The authors prove that any diagonal matrix D in (0,1)^N can be realized by a suitable choice of d (Theorem 2.1), derive analytical gradient expressions for the H2 approximation cost (Theorem 3.1), and provide upper and lower bounds on the multi-task approximation error (Theorems 4.1–4.3). Numerical experiments with aircraft, pendulum, and bicycle examples illustrate the approach.

Significance. The paper addresses a timely and interesting question: how a single static neural architecture with a low-dimensional task-dependent bias can approximate multiple linear controllers. Theorem 2.1 gives a clean and explicit parametrization of the linearized controller dynamics by the bias vector, and Theorem 3.1 provides usable gradient formulas. The upper and lower bounds, although partly technical, attempt to quantify how the approximation error depends on the number of tasks, the controller dimension, and the similarity of the desired systems, which is a useful step for this class of architectures. The numerical studies support the viability of the approach, though they do not by themselves establish the theoretical claims.

major comments (3)

[Section II, Eq. (5)] The H2 norm in problem (5) is finite only if the error system is stable; for stable desired systems A_i, this reduces to requiring that every linearized controller matrix -I + D_i W is Hurwitz. The paper does not restrict the optimization to this stable set, nor does it define the H2 norm as extended-valued (equal to +∞) for unstable error systems. Consequently, the gradient expressions in Theorem 3.1, which use the Gramians of the error system, are undefined for iterates that yield an unstable linearization, and gradient descent has no safeguard to remain in the well-defined region. This is load-bearing because the upper bound in Theorem 4.1 is expressed through the optimum of problem (5). Please add an explicit stability constraint or an extended-value definition of the cost, and describe how the proposed algorithm handles the stability boundary.
[Section IV-B, Theorem 4.3 proof] The proof of Theorem 4.3 explicitly omits the cases r ≤ r_j and r ≥ r_l, stating that they are 'not included' and lead to analogous conclusions. These cases are part of the theorem's stated domain, and the theorem claims a lower bound over the full problem. The omitted cases must be treated explicitly, or an argument must be provided showing that they are dominated by the included cases; otherwise the proof is incomplete and the theorem is not fully established.
[Section IV-A, Theorem 4.1 proof] The derivation of the upper bound is too terse at a critical point. The expression for J_B, which includes the term 2N ||Σ_aux||_∞, is introduced without explaining how it follows from the balanced-truncation error bound in reference [30]. The proof simply states that the bound is obtained by 'leveraging' the balanced truncation result. Since Theorem 4.1 is one of the main theoretical contributions, the proof should be expanded to show the origin of J_B and to specify which balanced-truncation result is used for the H2 norm.

minor comments (3)

[Remark 1] The parameter count for storing M distinct linear controllers is given as 'M n^4 m p', which appears to be a typo; the number of free parameters in a state-space realization (A, B, C) is n^2 + nm + pn (or n^2 + nm + pn + pm if a direct feedthrough term is included), not n^4 m p.
[Theorem 2.1 proof] The sentence 'Notice that φ_d is a injective function' should read 'an injective function'.
[Section IV-B, proof of Theorem 4.3] In the table after equation (26), the entry x_j = -∞ in Case 3 is not an admissible finite value; please clarify that this denotes a limiting value as x_j tends to -∞, and state the sense in which the infimum is attained.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the parameterization, gradient formulas, and bounds are derived from first principles and standard external results, with no fitted input renamed as a prediction and no load-bearing self-citation.

full rationale

The paper's derivation chain is self-contained. Theorem 2.1 explicitly constructs d and xeq for any desired diagonal D in (0,1)^N, and the claim that any linear controller can be replicated follows because W is free, so -I + D W can be set to any desired stable A. This is an exact parameterization, not a fitted prediction. The gradient formulas in Theorem 3.1 are derived by direct chain-rule computation using standard Lyapunov/Gramian identities, and they depend only on the problem's own variables. The upper bound in Theorem 4.1 is established by an explicit choice of W, D, B, and C plus a standard balanced-truncation error estimate from an external reference; it is not calibrated to numerical experiments, and the bound's comparison to the gradient-trained cost in Fig. 5 is an evaluation, not a fit. The lower bounds in Theorems 4.2 and 4.3 use elementary norm inequalities and the Lambert W function from an external, non-overlapping reference. The paper's self-citations (e.g., astrocyte motivation in [3], related multi-task identification in [15]) are contextual and not load-bearing for any theorem. The stability well-posedness issue noted by the reader, namely that problem (5) and the gradient formulas require each -I + D_i W to be Hurwitz while the paper imposes no explicit stability constraint, is a formal rigor gap rather than a circularity: the gradient and bounds are not derived by assuming the conclusion, and the paper explicitly acknowledges that the gradient procedure 'offers no stability or performance guarantees.' No step reduces by definition to its inputs, and no fitted quantity is presented as a prediction.

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

The theoretical claims rest on standard control-theoretic tools and explicit stability/domain assumptions. No ad hoc fitted parameters or invented physical entities are introduced.

assumptions (5)

domain assumption The desired systems Σ_i^D are stable, controllable, and observable.
Stated before equation (4); needed for finite H2 norms and for the balanced realization of the stacked system Σ_ext.
domain assumption The nonlinear neural controller operates near its bias-dependent equilibrium so that the linearized dynamics (3) accurately represent it.
The approximation error in (5) is defined on linearized systems, not on the nonlinear controller; no region-of-attraction or Lipschitz analysis is provided.
domain assumption Every linearized approximating system (-I + D_i W, B, C) is stable, so the H2 norm of the error system is finite.
Used implicitly in (5) and in the Gramian gradient formulas of Theorem 3.1; the paper does not enforce or verify stability.
standard math Balanced truncation error bounds from [30] and Lyapunov/H2 gradient results from [26] hold as cited.
Theorem 4.1 and Theorem 3.1 rely on these standard results.
standard math Properties of the Lambert W function W_{-1} used in Theorem 4.3 hold as cited from [31].
The scalar lower-bound proof depends on the negative branch of the Lambert W function.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Controllable Neural Architectures for Multi-Task Control." pith.science (2026). https://pith.science/paper/QORCLTIK

@misc{pith2026250119059,
  author       = {Pith},
  title        = {Pith review of: Controllable Neural Architectures for Multi-Task Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QORCLTIK}},
  note         = {Machine review of arXiv:2501.19059}
}

read the original abstract

This paper studies a multi-task control problem where multiple linear systems are to be regulated by a single non-linear controller. In particular, motivated by recent advances in multi-task learning and the design of brain-inspired architectures, we consider a neural controller with (smooth) ReLU activation function. The parameters of the controller are a connectivity matrix and a bias vector: although both parameters can be designed, the connectivity matrix is constant while the bias vector can be varied and is used to adapt the controller across different control tasks. The bias vector determines the equilibrium of the neural controller and, consequently, of its linearized dynamics. Our multi-task control strategy consists of designing the connectivity matrix and a set of bias vectors in a way that the linearized dynamics of the neural controller for the different bias vectors provide a good approximation of a set of desired controllers. We show that, by properly choosing the bias vector, the linearized dynamics of the neural controller can replicate the dynamics of any single, linear controller. Further, we design gradient-based algorithms to train the parameters of the neural controller, and we provide upper and lower bounds for the performance of our neural controller. Finally, we validate our results using different numerical examples.

Figures

Figures reproduced from arXiv: 2501.19059 by the authors.

**Figure 2.** This figure displays the impulse responses of the feedback inter [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** This figure illustrates a box plot summarizing the final cost (5) [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: This figure shows a box plot (with the format introduced in Fig. 3) [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: The figure illustrates two distinct curves: the red one represents [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 27 canonical work pages

[22]

L∞ norm simulta- neous system approximation,

D. Kavrano ˘glu, M. Bettayeb, and M. F. Anjum, “ L∞ norm simulta- neous system approximation,” vol. 6, no. 9-10, pp. 999–1014, 1996

work page 1996
[30]

Approximation of large-scale dynamical systems: an overview,

A. C. Antoulas and D. C. Sorensen, “Approximation of large-scale dynamical systems: an overview,” Int. J. Applied Mathematics and Computer Science, vol. 11, no. 5, pp. 1093–1121, 2001

work page 2001
[1]

A model of inductive bias learning,

J. Baxter, “A model of inductive bias learning,” Journal of Artificial Intelligence Research, vol. 12, pp. 149–198, 2000

work page 2000
[2]

Multitask learning,

R. Caruana, “Multitask learning,” Machine Learning , vol. 28, no. 1, p. 41, 1997

work page 1997
[3]

Astrocytes as a mechanism for meta-plasticity and contextually-guided network function,

L. Gong, F. Pasqualetti, T. Papouin, and S. Ching, “Astrocytes as a mechanism for meta-plasticity and contextually-guided network function,” 2024

work page 2024
[4]

Packnet: Adding multiple tasks to a single network by iterative pruning,

A. Mallya and S. Lazebnik, “Packnet: Adding multiple tasks to a single network by iterative pruning,” in IEEE Conf. on Computer Vision and Pattern Recognition, Los Alamitos, CA, USA, jun 2018, pp. 7765– 7773

work page 2018
[5]

A review of recurrent neural networks: Lstm cells and network architectures,

Y . Yu, X. Si, C. Hu, and J. Zhang, “A review of recurrent neural networks: Lstm cells and network architectures,” Neural computation, vol. 31, no. 7, pp. 1235–1270, 2019

2019
[6]

Neural networks for control and system identification,

P. J. Werbos, “Neural networks for control and system identification,” Tampa, FL, USA, Dec. 1989, pp. 260–265

work page 1989

Show all 31 references

[7]

B. D. Anderson and J. B. Moore, Optimal control: linear quadratic methods. Courier Corporation, 2007

2007
[8]

Model predictive control,

B. Kouvaritakis and M. Cannon, “Model predictive control,” Switzer- land: Springer International Publishing , vol. 38, pp. 13–56, 2016

2016
[9]

Overcoming the limitations of adaptive control by means of logic-based switching,

J. P. Hespanha, D. Liberzon, and A. S. Morse, “Overcoming the limitations of adaptive control by means of logic-based switching,” Systems & Control Letters , vol. 49, no. 1, pp. 49–65, 2003

2003
[10]

Adaptive control,

K. J. ˚Astr¨om, “Adaptive control,” in Mathematical System Theory: The Influence of R. E. Kalman , A. C. Antoulas, Ed. Springer Berlin Heidelberg, 1991, pp. 437–450

1991
[11]

The order of any stabilizing regulator is sufficient a priori information for adaptive stabilization,

B. M ˚artensson, “The order of any stabilizing regulator is sufficient a priori information for adaptive stabilization,” Systems & Control Letters, vol. 6, no. 2, pp. 87–91, 1985

1985
[12]

Adaptive stabilization of linear systems via switching control,

M. Fu and B. Barmish, “Adaptive stabilization of linear systems via switching control,” IEEE Transactions on Automatic Control , vol. 31, no. 12, pp. 1097–1103, 1986

1986
[13]

An adaptive controller which provides an arbitrarily good transient and steady-state response,

D. Miller and E. Davison, “An adaptive controller which provides an arbitrarily good transient and steady-state response,” IEEE Transac- tions on Automatic Control , vol. 36, no. 1, pp. 68–81, 1991

1991
[14]

Identifying the dynamics of a system by leveraging data from similar systems,

L. Xin, L. Ye, G. Chiu, and S. Sundaram, “Identifying the dynamics of a system by leveraging data from similar systems,” Atlanta, GA, USA, Jun. 2022, pp. 818–824

2022
[15]

Multi- task system identification of similar linear time-invariant dynamical systems,

Y . Chen, A. M. Ospina, F. Pasqualetti, and E. Dall’Anese, “Multi- task system identification of similar linear time-invariant dynamical systems,” Marina Bay Sands, Singapore, Dec. 2023, to appear. arXiv preprint arXiv:2301.01430

2023 arXiv
[16]

Multi-task imitation learning for linear dynamical systems,

T. T. Zhang, K. Kang, B. D. Lee, C. Tomlin, S. Levine, S. Tu, and N. Matni, “Multi-task imitation learning for linear dynamical systems,” arXiv preprint arXiv:2212.00186 , 2022

2022 arXiv
[17]

Imitation and transfer learning for LQG control,

T. Guo, A. A. Al Makdah, V . Krishnan, and F. Pasqualetti, “Imitation and transfer learning for LQG control,” vol. 7, pp. 2149–2154, 2023

2023
[18]

The application of artificial intelligence to biology and neuroscience,

B. Richards, D. Tsao, and A. Zador, “The application of artificial intelligence to biology and neuroscience,” Cell, vol. 185, no. 15, pp. 2640–2643, 2022

2022
[19]

Continual lifelong learning with neural networks: A review,

G. I. Parisi, R. Kemker, J. L. Part, C. Kanan, and S. Wermter, “Continual lifelong learning with neural networks: A review,” Neural Networks, vol. 113, pp. 54–71, 2019

2019
[20]

Multi-task linear bandits,

M. Soare, O. Alsharif, A. Lazaric, and J. Pineau, “Multi-task linear bandits,” in NIPS2014 Workshop on Transfer and Multi-task Learning: Theory meets Practice , 2014

2014
[21]

Multi-task learning for contextual bandits,

A. A. Deshmukh, U. Dogan, and C. Scott, “Multi-task learning for contextual bandits,” in Advances in Neural Information Processing Systems, vol. 30. Curran Associates, Inc., 2017, pp. 4851–4859

2017
[23]

Functional roles of astrocyte calcium elevations: From synapses to behavior,

S. Guerra-Gomes, N. Sousa, L. Pinto, and J. F. Oliveira, “Functional roles of astrocyte calcium elevations: From synapses to behavior,” Frontiers in Cellular Neuroscience , vol. 11, p. 427, 2018

2018
[24]

Astrocytes integrate behavioral state and vascular signals during functional hyperemia,

C. H. T. Tran, G. Peringod, and G. R. Gordon, “Astrocytes integrate behavioral state and vascular signals during functional hyperemia,” Neuron, vol. 100, no. 5, pp. 1133–1148.e3, 2018

2018
[25]

A conceptual framework for astrocyte function,

C. Murphy-Royal, S. Ching, and T. Papouin, “A conceptual framework for astrocyte function,” vol. 26, no. 11, pp. 1848–1856, 2023

2023
[26]

The smoothed spectral abscissa for robust stability opti- mization,

J. Vanbiervliet, B. Vandereycken, W. Michiels, S. Vandewalle, and M. Diehl, “The smoothed spectral abscissa for robust stability opti- mization,” vol. 20, no. 1, pp. 156–171, 2009

2009
[27]

K. B. Petersen and M. S. Pedersen, The Matrix Cookbook. Technical University of Denmark, 2012

2012
[28]

K. J. ˚Astr¨om and R. M. Murray, Feedback Systems, 2008

2008
[29]

Truncated balanced realization of a stable non-minimal state-space system,

M. S. Tombs and I. Postlethwaite, “Truncated balanced realization of a stable non-minimal state-space system,” International Journal of Control, vol. 46, pp. 1319–1330, 1987

1987
[31]

On the lambert w function,

R. M. Corless, G. H. Gonnet, D. E. G. Hare, D. J. Jeffrey, and D. E. Knuth, “On the lambert w function,” Advances in Computational Mathematics, vol. 5, no. 1, pp. 329–359, 1996

1996

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [22]

L∞ norm simulta- neous system approximation,

D. Kavrano ˘glu, M. Bettayeb, and M. F. Anjum, “ L∞ norm simulta- neous system approximation,” vol. 6, no. 9-10, pp. 999–1014, 1996

work page 1996

[2] [30]

Approximation of large-scale dynamical systems: an overview,

A. C. Antoulas and D. C. Sorensen, “Approximation of large-scale dynamical systems: an overview,” Int. J. Applied Mathematics and Computer Science, vol. 11, no. 5, pp. 1093–1121, 2001

work page 2001

[3] [1]

A model of inductive bias learning,

J. Baxter, “A model of inductive bias learning,” Journal of Artificial Intelligence Research, vol. 12, pp. 149–198, 2000

work page 2000

[4] [2]

Multitask learning,

R. Caruana, “Multitask learning,” Machine Learning , vol. 28, no. 1, p. 41, 1997

work page 1997

[5] [3]

Astrocytes as a mechanism for meta-plasticity and contextually-guided network function,

L. Gong, F. Pasqualetti, T. Papouin, and S. Ching, “Astrocytes as a mechanism for meta-plasticity and contextually-guided network function,” 2024

work page 2024

[6] [4]

Packnet: Adding multiple tasks to a single network by iterative pruning,

A. Mallya and S. Lazebnik, “Packnet: Adding multiple tasks to a single network by iterative pruning,” in IEEE Conf. on Computer Vision and Pattern Recognition, Los Alamitos, CA, USA, jun 2018, pp. 7765– 7773

work page 2018

[7] [5]

A review of recurrent neural networks: Lstm cells and network architectures,

Y . Yu, X. Si, C. Hu, and J. Zhang, “A review of recurrent neural networks: Lstm cells and network architectures,” Neural computation, vol. 31, no. 7, pp. 1235–1270, 2019

2019

[8] [6]

Neural networks for control and system identification,

P. J. Werbos, “Neural networks for control and system identification,” Tampa, FL, USA, Dec. 1989, pp. 260–265

work page 1989