acceptodds
Under review as a conference paper at ICLR 2027

Neural Real-Time Recurrent Learning

Abstract

Recurrent neural networks (RNNs) are commonly trained using backpropagation through time. However, its memory cost grows with sequence length, and it prevents online updates of the RNN parameters. Real-Time Recurrent Learning (RTRL) enables online parameter updates at a fixed memory cost. It maintains an influence matrix describing how parameter changes affect the state of the network. Despite these benefits, both its memory and per-step computational costs scale poorly with network width. Prior work has addressed these limitations either by restricting the recurrent architecture to make RTRL tractable or by approximating the influence matrix, notably through low-rank tensor representations. In this work, we introduce Neural Real-Time Recurrent Learning (NRTRL), which learns a compact representation of the influence matrix. NRTRL uses neural networks to update this representation over time and reconstruct the corresponding influence estimate, allowing the approximation mechanism itself to be learned. We instantiate this approach by reusing the RNN cell along a short virtual trajectory, with learned low-rank perturbations to its parameters and learned inputs. We show that this construction learns influence estimates that are accurate enough to enable learning. We evaluate NRTRL on pixel-by-pixel MNIST and character-level Penn Treebank.

Primary area:
learning on time series and dynamical systems
Keywords:
recurrent neural networks, real-time recurrent learning, online learning, temporal credit assignment, gradient approximation, memory-efficient training

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_mx66OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces NRTRL by framing the approximation of the RTRL influence matrix itself as a learnable problem. It adopts the unified perspective of encoder, decoder, and update operator proposed by Marschall et al. (2020): an influence code z(t)z(t) represents M(t)M(t), an update network UηU_\eta maintains this code, and an influence network decodes it into M^(t)\hat{M}(t). The key construction involves reusing the original RNN cell: learnable virtual inputs, virtual states, and low-rank parameter perturbations (W+URVW + URV) are applied over a virtual trajectory of length KK, and the Jacobian of the final virtual state with respect to parameters ww serves as the influence estimate (Eq. 6–8). This approach ensures the output is the Jacobian of a specific function (following the Sobolev approach) and maintains the structure of an influence matrix, all while relying solely on JVP/VJP operations without explicitly constructing the full matrix. Training employs a consistency loss to align "decoding the new code and taking one RTRL step" with the "RTRL target derived from the decoded old code" (Eq. 9); it uses Gaussian probes to express the Frobenius difference in a form suitable for unbiased estimation and applies bootstrapping with stop-gradient on the target (Eq. 10). Two theoretical results are presented: the contractivity of the RTRL operator TT under the RNN contraction assumption, and the fact that the average influence error is bounded by the average consistency loss (Proposition 1); and the formulation of the influence estimate as a sum of KK Kronecker products (Proposition 2, establishing the connection). Experiments conducted on pixel-by-pixel MNIST and character-level Penn Treebank compare the method against UORO, KF-RTRL, SnAp-1, and TBPTT variants (7/56 and 25), highlighting the advantage of a code size of O(H)O(H) compared to the O(H2)O(H^2) complexity of the baselines.

Critical strengths

Reframing the task from "designing a low-rank/sparse structure" to "learning an approximator" brings a clear, novel perspective to RTRL influence approximation. This approach bridges RTRL with Sobolev training and amortized optimization; furthermore, the mechanism relies exclusively on JVPs/VJPs—avoiding the full Dh×PD_h \times P matrix—making for a clean engineering implementation. Parameterizing the influence network using the RNN cell itself is highly parameter-efficient, while the combination of virtual trajectories and low-rank perturbations satisfies the dual constraints that the estimate must represent a Jacobian and must share the same architecture.

Additionally, the consistency loss used in the probe variant serves as an unbiased estimator of the Frobenius norm; the convergence analysis and the Kronecker-sum formulation in Proposition 2 effectively link the method to OK, demonstrating that the approach is not merely a haphazard assembly of ideas. The authors are also transparent about negative results—explicitly noting that SnAp-1 learns the fastest and best, that NRTRL performs only on par with KF-RTRL on MNIST, and that the method itself is biased—which makes the work more credible than most manuscripts on RTRL approximation.

Critical weaknesses

(1) The most clear-cut empirical value proposition lies in memory usage; however, this advantage relies on amortizing shared network parameters across the batch, and there is a complete lack of measured peak GPU memory data. The right-hand panels of Figures 3 and 4 do indeed show two bars: one for the "code" itself and another for "code + shared parameters/batch size 32" (in a lighter shade). Based on the MNIST (H=128) readings, the NRTRL code size is 3,845, rising to 8,631 when including the shared parameters divided by a batch size of 32; meanwhile, the smallest baseline code size is approximately 50,048. Thus, even when amortizing the shared parameters over a batch of 32, NRTRL remains roughly 5.8 times more efficient. The direction of this comparison is valid, and I do not consider the methodology "wrong." However, there are three genuine shortcomings. First, the shared parameters themselves are substantial: deducing from the difference (8,631 − 3,845 = 4,786), the update transformer (Table 2) and the U/V matrices of the influence network total around 150,000 scalar values—roughly 40 times the size of the "code" when H=128. Second, the amortization ratio depends entirely on the batch size. With a batch size of 1, the full 150,000 scalars must be accounted for, bringing the NRTRL total to approximately 157,000; this is significantly higher than the two baselines in the 50k range and comparable only to the largest baseline (approx. 164,000). While the paper does not claim that a batch size of 1 is mandatory, the targeted online use case could involve small batches; therefore, results across a range of batch sizes should at least have been reported. Third, there is no mention of Adam’s first- and second-order moments, activations, or peak memory usage in bytes. For a paper whose core claim centers on efficiency, providing only scalar code sizes—without curves showing batch-size dependency or actual measured memory usage—is unconvincing.

(2) "OK" does not appear at all in the main experiments. OK (Benzing et al., 2019) appears in the cost table in Figure 1(a), the theoretical discussion, and the detailed derivations in Appendix E.5—where it is treated as a stronger scheme within the Kronecker family—yet it is entirely absent from Section 6. The omission of OK represents a significant gap: it belongs to the same category as NRTRL (influence approximation), and Proposition 2 in this paper explicitly demonstrates that NRTRL can be formulated as a Kronecker sum. Consequently, without a comparison, it is impossible to answer the question of "whether a learned approximation outperforms an optimized fixed-structure approximation"—which is precisely the paper's core selling point. Regarding the approaches mentioned in related work that make RTRL feasible through architectural constraints (Irie et al. 2024; Elelimy et al. 2024), their reliance on different recurrent architectures makes them not directly comparable; however, the paper should have explicitly stated this scope limitation rather than implicitly assuming it covered the entire space of approximate RTRL methods.

(3) While the per-step computational complexity is O(H²), the actual constant factors may be substantial, yet the paper provides no wall-clock time measurements. Each step requires Nprobe=5 probes, with each probe involving three JVP passes over a virtual trajectory of length K=7 (for MNIST; K=5 for PTB), in addition to the online training updates for the transformer (Algorithm 4, lines 5–6, requires recomputing z(t) after a gradient step). Under the MNIST configuration, the dominant term in Eq. 189 is approximately 6×(7Cvirt + Cphys); ignoring CU, this equates to roughly 48 times the cost of a standard O(H²) recurrent/virtual step. This suggests a potentially large constant factor compared to UORO or SnAp-1, though without empirical speed comparisons, one cannot simply claim it is "dozens of times slower." From a fairness perspective: compared to the O(H³) complexity of KF-RTRL and the O(rH³) of OK, NRTRL actually holds an asymptotic advantage for large H; nevertheless, actual speed performance still requires verification via wall-clock timing. The paper glosses over efficiency with Big O notation; lacking FLOPs, ms/step, or wall-clock training times, it is impossible to assess actual efficiency.

(4) The theory holds only for strongly contracting RNNs, and the constants involved may render the bounds vacuous. The bound in Proposition 1 (Eq. 14) is general—applying to any M^\hat{M}—and Eq. 11 explicitly separates the decoder limit ϵM\epsilon_M from the update limit ϵU\epsilon_U. Thus, the paper directly addresses the question of whether targets generated by the previous step's estimates might converge to an incorrect fixed point; I do not see a logical flaw there. However, there are three genuine limitations. First, the assumption ∥J(t←s)∥2≤κρt−s\|J^{(t \leftarrow s)}\|_2 \leq \kappa \rho^{t-s} (ρ<1\rho < 1) implies global contraction, effectively excluding the very regimes where RNNs most need long-range memory. Second, the factor κ2/(1−ρ)2\kappa^2/(1-\rho)^2 blows up as ρ→1\rho \to 1: in practice, even if the consistency loss is minimized, the corresponding influence error bound may remain extremely loose; consequently, the paper fails to explain how effectively "consistency loss as a principled training objective" guides learning outside the strong-contraction regime. Third, Eq. 14 bounds the average influence error (Frobenius norm); linking this to the actual task gradient c(t)M^(t)c(t)\hat{M}(t) requires ∥c(t)∥\|c(t)\| to be bounded—a step that is not discussed.

(5) The comparison protocol conflates "update frequency" with "methodological superiority." Online methods update parameters at every step, whereas TBPTT-25 updates only every 25 steps; the main text itself acknowledges that "online methods make many more updates for the same data." The MNIST experiments use only three random seeds and lack significance testing, while the curves for NRTRL and SnAp-1 on the PTB dataset are indistinguishable. Given the constant-factor overhead associated with NRTRL, this comparison hardly supports a conclusion of "higher efficiency."

(6) The update protocols differ across baselines; this discrepancy should be noted in the conclusions. Appendix I clarifies that SnAp-1 updates its sparse influence estimate at every physical step, whereas NRTRL, KF-RTRL, and UORO update only once per TBPTT window (treating the sequence of LL steps as a single transition). This results in a higher update frequency for SnAp-1, which may give it an advantage in comparisons. While the paper discloses that all methods utilize TBPTT windows, the issue lies not in whether a method is "sufficiently online," but in the differing update frequencies of the carried representations across baselines. This point should be explicitly stated in the interpretation of results to prevent readers from misconstruing differences in update frequency as differences in approximation quality.

The abstract states that the code size is "several orders of magnitude" smaller than the full influence matrix. While this statement is technically correct (the full matrix is ​​on the order of H3H^3, while the NRTRL code is around 3.8k–5.7k—a difference of 2–3 orders of magnitude), it is easily misinterpreted as an advantage relative to the baselines. The actual difference compared to the baselines is approximately 13–43× for MNIST and 41–114× for PTB—representing a difference of 1–2 orders of magnitude, not "several." It is recommended to explicitly state the denominator in the main text.

Questions for authors
  • Once θ, η, and Adam momentum terms are all accounted for, what is the per-sequence/total peak memory usage (in bytes)? What proportion of the total memory is consumed by the O(H) code? Is it possible to provide wall-clock time and FLOPs comparisons against SnAp-1, KF-RTRL, and UORO on the same hardware?
  • Why was OK not included in the experiments? How do NRTRL and OK compare in terms of accuracy, memory usage, and runtime under identical H and training budget constraints?
  • Under what ranges of κ and ρ is the bound in Eq. 14 tight? To ensure the task gradient error remains controllable, to what level must the consistency loss be reduced? In the experiments actually conducted (GRU, MNIST/PTB), what were the measured values ​​for κ, ρ, and the average consistency loss?
  • Could you show the relationship between the task gradient error ‖c(t)M̂(t) − c(t)M(t)‖ and the consistency loss, rather than just the influence error? How tight is this bound in practice given the variance of the Gaussian probe?
  • Can the comparison with TBPTT be performed on an update-budget-matched basis (where each method is allowed the same number of parameter updates)? Otherwise, is the "higher accuracy" simply a result of performing more updates?
2 comments

We thank you for your thorough review. We cover your questions about memory, the constant factor and the update protocols in the general reply, and your questions about OK, the theory, update frequency and code size below.

OK. We acknowledge that Proposition 2 writes NRTRL's estimate as a sum of KK Kronecker products, which connects it to OK. This connection is in the general form of the approximation, and the two methods are very different. OK stores the Kronecker factors, whereas NRTRL generates them from its code through the influence network. With rr Kronecker products, OK needs rr times the memory of KF-RTRL, which already grows quadratically with the width. We left OK out because KF-RTRL already represents this family with less memory, and OK adds implementation complexity.

A future version will clarify that this work is only about the approximations of RTRL, and that we do not want to compare to methods that make RTRL tractable.

Scope of the theory. We agree that the contraction assumption excludes long memory, and that the factor κ2/(1−ρ)2\kappa^2/(1-\rho)^2 grows as ρ\rho approaches 1. Proposition 1 presents one setting in which the consistency loss is a principled objective, and does not claim that its assumption holds in our experiments. We believe a result under limiting assumptions is still more informative than none. A future version will state these limits explicitly, and add the step from influence error to gradient error,

∥c(t)(M^(t)−M(t))∥≤∥c(t)∥∥M^(t)−M(t)∥F,\lVert \mathbf{c}^{(t)} (\widehat{\mathbf{M}}^{(t)} - \mathbf{M}^{(t)}) \rVert \le \lVert \mathbf{c}^{(t)} \rVert \lVert \widehat{\mathbf{M}}^{(t)} - \mathbf{M}^{(t)} \rVert_F,

which bounds the gradient error when the credit c(t)\mathbf{c}^{(t)} is bounded.

Update frequency. Frequent updates are the purpose of online methods, and the works we compare with use the same protocol [1, 2, 3]. Thus we believe this is not a flaw of the protocol. It shows why online update methods like ours are relevant against TBPTT. A future version will also include TBPTT over one step, which updates the parameters as often as the online methods.

Code size against the baselines. We agree. A future version will state that "several orders of magnitude" refers to the full influence matrix, and give the ratio to each baseline separately.

References

[1] A. Mujika, F. Meier and A. Steger. Approximating Real-Time Recurrent Learning with Random Kronecker Factors. NeurIPS, 2018.

[2] F. Benzing, M. M. Gauy, A. Mujika, A. Martinsson and A. Steger. Optimal Kronecker-Sum Approximation of Real Time Recurrent Learning. ICML, 2019.

[3] J. Menick, E. Elsen, U. Evci, S. Osindero, K. Simonyan and A. Graves. Practical Real Time Recurrent Learning with a Sparse Approximation. ICLR, 2021.

Thanks for your clarification and explanation, which help me better understand your design and claims.

Reviewer_KNG2OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors introduce Neural Real-Time Recurrent Learning (NRTRL), an approach to approximating the influence matrix in Real-Time Recurrent Learning (RTRL). Although RTRL enables online learning without storing past hidden states, its prohibitive memory and computational costs limit its practical applicability. Existing approximation methods address these limitations through stochastic low-rank factorizations or sparse representations of the influence matrix. In contrast, NRTRL learns to predict a compact influence representation using neural networks, allowing the approximation mechanism itself to be learned rather than predefined.

Specifically, NRTRL employs an influence network that constructs influence estimates as the Jacobian of a short virtual recurrent trajectory, using learned inputs and low-rank parameter perturbations. An additional update network learns to propagate the compact influence representation over time by minimizing a consistency loss derived from the RTRL recurrence. This approach reduces the theoretical per-sequence memory complexity of RTRL influence matrix approximations from O(H2)O(H^2) to O(H)O(H), while maintaining O(H2)O(H^2) computational complexity per step for fixed hyperparameters. The memory comparison excludes neural network parameters shared across sequences.

The authors evaluate NRTRL on sequential MNIST and character-level Penn Treebank, comparing it against existing RTRL approximations, including UORO, KF-RTRL, and SnAp-1, as well as truncated backpropagation through time (TBPTT). While NRTRL achieves reasonable task performance with substantially smaller influence representations, it does not consistently outperform existing methods, particularly SnAp-1, which demonstrates better performance in both experiments. Moreover, the experiments primarily evaluate NRTRL in combination with TBPTT, making it difficult to isolate the effectiveness of the proposed influence approximation. The paper also lacks direct evaluations of influence estimation accuracy and empirical computational efficiency, leaving the practical advantages of NRTRL beyond memory reduction insufficiently demonstrated.

Critical strengths

The proposed method reduces the theoretical per-sequence memory complexity of RTRL influence matrix approximations from O(H2)O(H^2) to O(H)O(H) by leveraging a compact representation based on a virtual recurrent trajectory. The experimental results further demonstrate a substantial reduction in influence code size in practical settings.

Critical weaknesses
  • Insufficient empirical validation of the proposed components and practical efficiency: The paper lacks component-wise ablation studies to justify the necessity and effectiveness of its proposed design choices, such as the virtual recurrent trajectory, learned inputs, and low-rank parameter perturbations. Furthermore, although the authors demonstrate a substantial reduction in influence code size and provide theoretical complexity analysis, these results do not directly establish improvements in practical memory consumption or computational efficiency. Empirical comparisons of peak GPU memory usage, FLOPs, and wall-clock training time would strengthen the paper's efficiency claims.

  • Unclear performance–efficiency trade-off compared to existing methods: Although NRTRL substantially reduces the influence code size, SnAp-1 achieves better task performance on both benchmarks. The smaller representation could justify this performance gap if it leads to meaningful improvements in practical efficiency. However, without comparisons under matched memory or computational budgets, it remains unclear whether NRTRL offers a favorable performance–efficiency trade-off over existing methods.

  • Lack of direct evaluation of influence approximation quality: The primary objective of NRTRL is to learn a compact approximation of the RTRL influence matrix. However, the experiments evaluate the method mainly through downstream task performance, without directly measuring the accuracy of the estimated influence matrices or gradients.

Questions for authors
  • (major) Q1. Component-wise ablation studies: Could the authors provide ablation studies to clarify the individual contributions of the proposed components, including the virtual recurrent trajectory, learned inputs, low-rank parameter perturbations, and update network? Additionally, how sensitive is the performance to the virtual trajectory length KK and perturbation rank rr?

  • (major) Q2. Practical memory and computational efficiency: Could the authors provide empirical comparisons of peak GPU memory usage, FLOPs, and wall-clock training time against existing RTRL approximations? Although NRTRL reduces the theoretical per-sequence memory complexity, it introduces additional neural networks and multiple Jacobian-vector products. It would be valuable to demonstrate whether the smaller influence representation translates into practical efficiency improvements.

  • (major) Q3. Performance–efficiency trade-off: Given that SnAp-1 outperforms NRTRL on both benchmarks, could the authors provide comparisons under matched memory or computational budgets? Such comparisons would help clarify whether NRTRL's reduced influence code size offers sufficient practical benefits to justify the performance gap.

  • (major) Q4. Accuracy of influence approximation: Could the authors directly evaluate the accuracy of the learned influence approximation by comparing it against exact RTRL on small-scale models? For example, reporting the relative influence matrix error or gradient cosine similarity across different sequence lengths would help establish whether NRTRL accurately captures long-term parameter sensitivities.

  • (minor) Q5. Combination with TBPTT: Why do the authors combine NRTRL with TBPTT in the experiments? Although the TBPTT-only baselines help assess the additional benefits of influence approximation, the current experimental setting does not establish how effectively NRTRL performs without multi-step BPTT. Could the authors provide results with a truncation window of L=1L=1 and investigate the sensitivity to the TBPTT window length?

1 comment

We thank you for your detailed questions. We cover your questions about ablations, measured efficiency, the accuracy of the influence estimate and the combination with TBPTT in the general reply, and your question about matched budgets below.

Matched budgets. Matched memory or compute budgets are unusual in this literature, where methods are compared on the same network [1, 2, 3]. We will consider a way to make this possible and useful in a future version.

References

[1] C. Tallec and Y. Ollivier. Unbiased Online Recurrent Optimization. ICLR, 2018.

[2] A. Mujika, F. Meier and A. Steger. Approximating Real-Time Recurrent Learning with Random Kronecker Factors. NeurIPS, 2018.

[3] F. Benzing, M. M. Gauy, A. Mujika, A. Martinsson and A. Steger. Optimal Kronecker-Sum Approximation of Real Time Recurrent Learning. ICML, 2019.

Reviewer_tfaWOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

For training RNNs in RTRL framework, this paper proposes Neural RTRL that represents the influence matrix as the Jacobian of a learned function. This function reused the recurrent cell with learnable virtual state, virtual input, and low-rank weight perturbations. The influence network and the update network are jointly optimized with the proposed consistency loss. The proposed method reduces the memory and computation cost from O(H^3) and O(H^4) to O(H) and O(H^2), respectively. The paper also provides theoretical analysis of the error decomposition, the influence estimate convergence, and the influence estimate decomposition. The proposed method is evaluated on pixel-by-pixel MNIST and character-level Penn Treebank benchmark with the GRU architectures.

Critical strengths
  • The proposed idea of the influence matrix approximation is interesting and novel. Representing the influence matrix as the Jacobian of a learnable function is well-motivated.
  • The memory cost scales well O(H)\mathcal{O}(H) compared with other baselines O(H2)\mathcal{O}(H^2)
  • This paper is self-contained and easy to follow. The proposed method is supported by a solid theoretical analysis.
Critical weaknesses

Whie the method is well motivated and supported by theoretical analysis, the experimental validation is relatively limited.

  • The main advantage of NRTRL is memory and computational cost reduction. However, it is difficult to confirm this effectiveness from the experimental results. It would be better to include actual GPU memory and wall-clock of the proposed method compared with other baselines. It would be best when the accuracy and BPC of the proposed method are similar with or better than SnAp-1 while the memory and computation cost are reduced.
  • There is not enough ablation study. The trajectory length, hidden dimension, and RNN cell size are needed to be ablated for support the effectiveness of the proposed method.
Questions for authors

Please refer the weakness

2 comments

We thank you for your positive assessment of the idea and the theory. We cover your questions about memory, wall-clock time and ablations in the general reply. A future version will also vary the hidden size.

Thank you for the clarification. Figure 1(a) highlights the favorable memory and per-step computational scaling of NRTRL relative to several existing methods. I therefore consider empirical evaluation of these costs important for assessing the significance of this contribution. My concern is not that NRTRL must outperform every baseline at every model size, but that the current experiments do not sufficiently establish how the theoretical advantages translate into actual training costs and trade-offs. As the cost measurements and ablation studies remain future work, my original concerns remain unresolved, and I maintain my weak reject recommendation.

Comments on the paper

2 comments

Dear Reviewers,

I'm writing to gently remind you that the rebuttal discussion phase deadline is approaching. Please take a moment to read the authors' rebuttal, and feel free to respond if any concerns remain that you'd like them to address. Thank you very much for your time and effort.

AuthorsOpenReview ↗

General reply

We thank the reviewers for taking the time to read our work and for their comments. Four points came up in several reviews which we answer here in a shared reply. The specific answers can be found below.

Costs and scale. We do not claim NRTRL uses fewer FLOPs or less peak memory at the sizes used in the submission. We claim that the memory each sequence carries grows linearly with the width of the RNN, against quadratically for the other approximations. This advantage is not obvious at the sizes we tried in this submission. We consider this work as a first step before scaling up. A future version will report the peak memory, FLOPs and time per step of every method at the sizes of the submission and across batch sizes, and the peak memory and FLOPs of much larger RNNs.

Training protocol. On MNIST, we combined the influence approximations with TBPTT over windows of 7 steps, as in memory-TT UORO [1]. This was a mistake. It blurs the contribution of the influence estimate. In a future version, every influence method will compute the gradient from its estimate alone and update its representation at every step.

Accuracy of the influence estimate. We agree that task performance does not measure the accuracy of an estimate. During training, a future version will compare the gradient of every method with the exact RTRL gradient, as [2, 3] did.

Design and ablations. The influence network reuses the recurrent cell because the direct alternatives, which we mention in Section 4, do not scale. A future version will test each component by removing the virtual inputs, removing the low-rank perturbations and by varying the length of the virtual trajectory. However, the update network cannot be removed, since the code would then have to be found by iterative optimization at every step (paragraph "Amortized solution").

References

[1] C. Tallec and Y. Ollivier. Unbiased Online Recurrent Optimization. ICLR, 2018.

[2] A. Mujika, F. Meier and A. Steger. Approximating Real-Time Recurrent Learning with Random Kronecker Factors. NeurIPS, 2018.

[3] F. Benzing, M. M. Gauy, A. Mujika, A. Martinsson and A. Steger. Optimal Kronecker-Sum Approximation of Real Time Recurrent Learning. ICML, 2019.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.