arxiv: 2605.11855 · v1 · submitted 2026-05-12 · 💻 cs.LG · cs.AI· cs.AR

Recognition: no theorem link

Improving the Performance and Learning Stability of Parallelizable RNNs Designed for Ultra-Low Power Applications

Julien Brandoit , Arthur Fyon , Damien Ernst , Guillaume Drion

Authors on Pith no claims yet

Pith reviewed 2026-05-13 06:28 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.AR

keywords RNNCMRUcumulative updategradient flowbistable memoryultra-low poweranalog hardwarelong-range dependencies

0 comments

The pith

The cumulative update formulation in CMRU and αCMRU restores gradient flow while preserving bistable memory for ultra-low power RNNs.

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

The paper identifies gradient blocking during state updates as the core limitation in the original Bistable Memory Recurrent Unit, which otherwise provides quantized states with hysteresis for persistent memory and direct analog hardware mapping. Replacing the update with a cumulative version sums increments over time steps, forming skip connections that allow gradients to propagate freely during training. This change yields the CMRU and its relaxed αCMRU variant, which train far more stably and show less sensitivity to random initialization. Experiments across benchmarks demonstrate that these units match or exceed the performance of Linear Recurrent Units and minimal Gated Recurrent Units at small scales, with the strongest gains on tasks needing discrete information retained over long distances. A reader would care because the approach keeps the hardware-efficient properties intact while solving the learning problems that previously limited practical use of such low-power designs.

Core claim

The paper establishes that introducing a cumulative formulation for state updates in the Bistable Memory Recurrent Unit creates the Cumulative Memory Recurrent Unit (CMRU) and its αCMRU variant, which restore gradient flow through time via skip connections, improve convergence stability and initialization robustness, and achieve competitive or superior results to LRUs and minGRUs on benchmarks while retaining the quantized states, persistent memory, and noise-resilient dynamics required for analog implementation.

What carries the argument

The cumulative update formulation that accumulates increments to produce skip-connections through time in the bistable recurrent dynamics.

If this is right

The cumulative formulation dramatically improves convergence stability and reduces initialization sensitivity.
CMRU and αCMRU match or outperform LRUs and minGRUs across diverse benchmarks at small model sizes.
Advantages are pronounced on tasks requiring discrete long-range retention.
CMRU retains quantized states, persistent memory, and noise-resilient dynamics essential for analog implementation.

Where Pith is reading between the lines

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

This method could apply to other recurrent units to enhance trainability under hardware constraints.
Direct analog mapping of the preserved dynamics might enable ultra-low power chips for real-time sequence processing.
Stability gains may support training on longer sequences or with fewer examples.
Extensions to hybrid digital-analog systems could be tested for further power reductions.

Load-bearing premise

That the cumulative update formulation preserves the quantized states with hysteresis, persistent memory, and noise-resilient dynamics required for direct analog hardware mapping while restoring gradient flow.

What would settle it

A training run on a long-range dependency benchmark where the CMRU shows no improvement in retention accuracy over long distances or exhibits vanishing gradients during backpropagation would falsify the central claim.

Figures

Figures reproduced from arXiv: 2605.11855 by Arthur Fyon, Damien Ernst, Guillaume Drion, Julien Brandoit.

**Figure 1.** Figure 1: Fixed point structure of hidden state ht for CMRU as a function of ε. Left (ε = 0): Original BMRU with two discrete stable fixed points at h ∗ = ±α. Right (ε = 1): CMRU cumulative formulation expands reachable fixed points to all integer multiples h ∗ ∈ {±kα : k ∈ N}, creating countably infinite stable attractors. Dashed vertical lines at ±β indicate candidate state thresholds for gating. 4 [PITH_FULL_IMA… view at source ↗

**Figure 2.** Figure 2: Performance on sequential MNIST (sMNIST) classification task. Results show mean accuracy across five random initializations with min–max error bars. Left: Effect of hyperparameter ε ∈ {0.00, 0.25, 0.50, 0.75, 1.00} on CMRU and αCMRU (d = 32, single layer, last pooling). CMRU with ε = 0 corresponds to BMRU. Increasing ε consistently improves mean accuracy and reduces variability. Right: Scaling across state… view at source ↗

**Figure 3.** Figure 3: MAE on copy-first-input (continuous, noisy) versus sequence length L. Left: r = 1, d = 4. Middle: r = 1, d = 16. Right: r = 4, d = 16. Dashed red lines: CMRU quantization limits E ∗ MAE. Results shown as mean with min–max error bars over five seeds. CMRU with ε = 1 maintains consistent performance across sequence lengths L and capacity configurations ( [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: Native CMRU analog circuit. (A) Block diagram of the conceptual architecture. The CMRU is realized as a cascade of BMRU cells (Fyon et al., 2026a) interleaved with Differential Pair Integrator (DPI) blocks. Each BMRU stage occupies one level of the fixed-point ladder, and each DPI integrates the output of the preceding stage toward a multiple of α, shifting the hysteresis window of the next stage. The fini… view at source ↗

**Figure 5.** Figure 5: Common backbone architecture used across all experiments. The architecture interleaves recurrent layers with MLPs, skip connections, and normalization. We study different recurrent cell types (CMRU, αCMRU, LRU, minGRU) by varying the internal mechanism within Cell( g ·) while keeping the backbone structure fixed. The model dimension is fixed at m = 256 across all experiments. The state dimension d refers… view at source ↗

**Figure 6.** Figure 6: Accuracy on the pixel-by-pixel MNIST classification task across different pixel orderings. Each architecture is evaluated over five random initializations (shown as scattered points), with results reported as mean across seeds with min–max error bars. Cell types are presented from left to right within each task as: CMRU with ε ∈ {0.00, 0.10, 0.25, 0.50, 0.75, 1.00}, then αCMRU for the same ε values, follow… view at source ↗

**Figure 7.** Figure 7: Performance on copy first input tasks as a function of sequence length L. Each architecture is evaluated over five random initializations (shown as scattered points), with results reported as mean across seeds with min–max error bars. Cell types are presented from left to right within each sequence length as: CMRU, αCMRU (both with ε = 1), LRU, and minGRU. All models use hidden state dimension d = 4 and a… view at source ↗

read the original abstract

Sequence learning is dominated by Transformers and parallelizable recurrent neural networks (RNNs) such as state-space models, yet learning long-term dependencies remains challenging, and state-of-the-art designs trade power consumption for performance. The Bistable Memory Recurrent Unit (BMRU) was introduced to enable hardware-software co-design of ultra-low power RNNs: quantized states with hysteresis provide persistent memory while mapping directly to analog primitives. However, BMRU performance lags behind parallelizable RNNs on complex sequential tasks. In this paper, we identify gradient blocking during state updates as a key limitation and propose a cumulative update formulation that restores gradient flow while preserving persistent memory, creating skip-connections through time. This leads to the Cumulative Memory Recurrent Unit (CMRU) and its relaxed variant, the $\alpha$CMRU. Experiments show that the cumulative formulation dramatically improves convergence stability and reduces initialization sensitivity. The CMRU and $\alpha$CMRU match or outperform Linear Recurrent Units (LRUs) and minimal Gated Recurrent Units (minGRUs) across diverse benchmarks at small model sizes, with particular advantages on tasks requiring discrete long-range retention, while the CMRU retains quantized states, persistent memory, and noise-resilient dynamics essential for analog implementation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

CMRU adds cumulative updates to fix BMRU gradient blocking and improve stability, but the analog hardware invariance claims lack direct verification.

read the letter

The core contribution is a cumulative state update that turns the BMRU recurrence into one with skip connections through time. This restores gradient flow while the authors claim the quantized bistable states and persistent memory stay intact, producing CMRU and the relaxed αCMRU variant. On the digital side the results look useful: the new units converge more reliably, show less initialization sensitivity, and match or exceed LRU and minGRU on small-model benchmarks, with clearer gains on tasks that need discrete long-range retention. That is a practical step forward for anyone trying to keep RNNs competitive at very low power. The main weakness is that the hardware co-design story rests on an unshown invariance. The abstract asserts that quantized states, hysteresis, and noise resilience survive the change, yet no fixed-point analysis, phase-plane check, or noise-injection experiment is referenced to confirm the bistable loop width and analog mapping properties remain the same. If the cumulative term shifts the equilibria even modestly, the claimed direct mapping to analog primitives disappears even if the floating-point benchmarks improve. The reported experiments also omit error bars, exact dataset splits, and ablations on the cumulative parameter, so the size of the stability gain is still hard to judge. This paper is aimed at researchers working on hardware-aware sequence models and ultra-low-power edge inference. A reader already building or evaluating parallel RNNs for constrained devices would get concrete value from the new recurrence and the benchmark numbers. I would send it to peer review; the idea is clean, the performance claims are testable, and the missing dynamical checks are straightforward to add.

Referee Report

1 major / 2 minor

Summary. The paper introduces the Cumulative Memory Recurrent Unit (CMRU) and its relaxed variant αCMRU as modifications to the Bistable Memory Recurrent Unit (BMRU). It identifies gradient blocking in BMRU state updates as a limitation and proposes a cumulative formulation that creates skip connections through time to restore gradient flow. The central claims are dramatically improved convergence stability, reduced initialization sensitivity, competitive or superior performance versus LRUs and minGRUs on diverse benchmarks at small sizes (especially discrete long-range retention tasks), and retention of quantized states, hysteresis, persistent memory, and noise resilience for direct analog hardware mapping.

Significance. If the invariance of the bistable properties under the cumulative update is established, the work would meaningfully advance hardware-software co-design for ultra-low-power sequence models by delivering improved digital-task performance without sacrificing the analog-primitive mapping advantages of BMRU. This addresses a key power-performance tradeoff in edge-deployed RNNs.

major comments (1)

[Abstract] Abstract: the claim that CMRU 'retains quantized states, persistent memory, and noise-resilient dynamics essential for analog implementation' is load-bearing for the hardware co-design motivation, yet the manuscript supplies neither a fixed-point analysis nor a derivation showing that the cumulative update leaves the hysteresis loop, bistable equilibria, and noise tolerance unchanged from BMRU.

minor comments (2)

[Abstract] Abstract and experiments: no error bars, dataset sizes, or ablation isolating the cumulative update are reported, which weakens verification of the 'dramatically improves convergence stability' claim.
The manuscript should clarify the precise recurrence equation for the cumulative update (including how the skip connection is implemented) to allow reproduction of the gradient-flow restoration.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address the concern regarding the preservation of bistable properties by adding the requested analysis in the revised manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: the claim that CMRU 'retains quantized states, persistent memory, and noise-resilient dynamics essential for analog implementation' is load-bearing for the hardware co-design motivation, yet the manuscript supplies neither a fixed-point analysis nor a derivation showing that the cumulative update leaves the hysteresis loop, bistable equilibria, and noise tolerance unchanged from BMRU.

Authors: We agree that an explicit fixed-point analysis is needed to rigorously support the claim. In the revised manuscript we will add a derivation (as a new subsection in Section 3) showing that the CMRU equilibria coincide exactly with those of the BMRU: the cumulative formulation can be rewritten as a state-dependent skip connection that does not alter the quantization thresholds or the locations of the stable fixed points, thereby leaving the hysteresis loop and noise-resilience properties unchanged. This analysis directly substantiates the abstract claim while preserving the hardware-mapping advantages. revision: yes

Circularity Check

0 steps flagged

No significant circularity; claims rest on external benchmarks and stated assumptions

full rationale

The paper identifies gradient blocking in BMRU as a limitation, proposes cumulative update to restore flow while claiming preservation of quantized states/hysteresis, and validates via digital benchmarks against LRU/minGRU. No equations reduce a prediction to a fitted input by construction, no load-bearing self-citation chain, and no ansatz or uniqueness theorem imported from prior self-work. Performance results are independent of the preservation claim, which is presented as a design property rather than a derived result equivalent to the input. This is the common case of a self-contained proposal with external validation.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

The work rests on standard RNN training assumptions and introduces new model entities; no explicit free parameters are described in the abstract.

axioms (1)

standard math Standard backpropagation through time enables gradient-based training of recurrent models
Invoked implicitly when identifying gradient blocking as the core limitation.

invented entities (2)

CMRU no independent evidence
purpose: RNN unit using cumulative updates for restored gradient flow while retaining bistable memory
Newly defined in this work to address BMRU limitations.
αCMRU no independent evidence
purpose: Relaxed variant of CMRU for additional flexibility
Newly defined in this work.

pith-pipeline@v0.9.0 · 5530 in / 1246 out tokens · 46044 ms · 2026-05-13T06:28:40.442295+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

64 extracted references · 64 canonical work pages

[1]

2024 , url=

Albert Gu and Tri Dao , booktitle=. 2024 , url=

work page 2024
[2]

International Conference on Learning Representations , year=

Efficiently Modeling Long Sequences with Structured State Spaces , author=. International Conference on Learning Representations , year=

work page
[3]

Transformers are

Tri Dao and Albert Gu , booktitle=. Transformers are. 2024 , organization=

work page 2024
[4]

International Conference on Machine Learning , pages=

Resurrecting Recurrent Neural Networks for Long Sequences , author=. International Conference on Machine Learning , pages=. 2023 , organization=

work page 2023
[5]

Advances in Neural Information Processing Systems , volume=

Attention Is All You Need , author=. Advances in Neural Information Processing Systems , volume=

work page
[6]

and Neuhoff, David L

Gray, Robert M. and Neuhoff, David L. , journal=. Quantization , year=

work page
[7]

International Conference on Learning Representations , year=

Decoupled Weight Decay Regularization , author=. International Conference on Learning Representations , year=

work page
[8]

International Conference on Machine Learning , pages=

Language Modeling with Gated Convolutional Networks , author=. International Conference on Machine Learning , pages=. 2017 , organization=

work page 2017
[9]

2016 , eprint=

Layer Normalization , author=. 2016 , eprint=

work page 2016
[10]

1991 , type=

Hochreiter, Sepp , title=. 1991 , type=

work page 1991
[11]

IEEE Transactions on Neural Networks , volume=

Bengio, Yoshua and Simard, Patrice and Frasconi, Paolo , title=. IEEE Transactions on Neural Networks , volume=. 1994 , doi=

work page 1994
[12]

2015 , eprint=

A Simple Way to Initialize Recurrent Networks of Rectified Linear Units , author=. 2015 , eprint=

work page 2015
[13]

Albert Gu and Tri Dao and Stefano Ermon and Atri Rudra and Christopher R\'e , booktitle=

work page
[14]

International Conference on Machine Learning , pages=

On the difficulty of training Recurrent Neural Networks , author=. International Conference on Machine Learning , pages=. 2013 , organization=

work page 2013
[15]

Advances in Neural Information Processing Systems , editor=

Voelker, Aaron and Kaji\'. Advances in Neural Information Processing Systems , editor=

work page
[16]

International Conference on Machine Learning , pages=

Unitary Evolution Recurrent Neural Networks , author=. International Conference on Machine Learning , pages=. 2016 , organization=

work page 2016
[17]

Advances in Neural Information Processing Systems , volume=

Full-Capacity Unitary Recurrent Neural Networks , author=. Advances in Neural Information Processing Systems , volume=

work page
[18]

Orthogonal Recurrent Neural Networks with Scaled

Kyle Helfrich and Devin Willmott and Qiang Ye , booktitle=. Orthogonal Recurrent Neural Networks with Scaled. 2018 , organization=

work page 2018
[19]

International Conference on Machine Learning , pages=

On orthogonality and learning recurrent networks with long term dependencies , author=. International Conference on Machine Learning , pages=. 2017 , organization=

work page 2017
[20]

, journal=

Boyd, Stephen and Chua, Leon O. , journal=. Fading memory and the problem of approximating nonlinear operators with. 1985 , publisher=

work page 1985
[21]

International Conference on Learning Representations , year=

Efficiently modeling long sequences with structured state spaces , author=. International Conference on Learning Representations , year=

work page
[22]

Advances in Neural Information Processing Systems , volume=

On the parameterization and initialization of diagonal state space models , author=. Advances in Neural Information Processing Systems , volume=

work page
[23]

International Conference on Machine Learning , year=

The illusion of state in state-space models , author=. International Conference on Machine Learning , year=

work page
[24]

Talathi and Aniket Vartak , year=

Sachin S. Talathi and Aniket Vartak , year=. Improving performance of recurrent neural network with. 1511.03771 , archivePrefix=

work page arXiv
[25]

International Conference on Machine Learning , pages=

Cheap orthogonality constraints in neural networks: A simple parametrization of the orthogonal and unitary group , author=. International Conference on Machine Learning , pages=. 2019 , organization=

work page 2019
[26]

Neural Computation , volume=

Long short-term memory , author=. Neural Computation , volume=. 1997 , publisher=

work page 1997
[27]

Learning phrase representations using

Cho, Kyunghyun and Van Merri. Learning phrase representations using. Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

work page 2014
[28]

Leo Feng and Frederick Tung and Hossein Hajimirsadeghi and Mohamed Osama and Yoshua Bengio and Roland Memisevic , year=. Were. 2410.01201 , archivePrefix=

work page arXiv
[29]

IEEE Signal Processing Magazine , volume=

Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks , author=. IEEE Signal Processing Magazine , volume=. 2019 , publisher=

work page 2019
[30]

2005 , publisher=

Theoretical neuroscience: computational and mathematical modeling of neural systems , author=. 2005 , publisher=

work page 2005
[31]

International Conference on Learning Representations , year=

Simplified state space layers for sequence modeling , author=. International Conference on Learning Representations , year=

work page
[32]

PLoS ONE , volume=

A bio-inspired bistable recurrent cell allows for long-lasting memory , author=. PLoS ONE , volume=. 2021 , publisher=

work page 2021
[33]

Neural Networks , volume=

Warming up recurrent neural networks to maximise reachable multistability greatly improves learning , author=. Neural Networks , volume=. 2023 , publisher=

work page 2023
[34]

International Conference on Learning Representations , year=

Long Range Arena: A Benchmark for Efficient Transformers , author=. International Conference on Learning Representations , year=

work page
[35]

2019 , eprint=

Overcoming the vanishing gradient problem in plain recurrent networks , author=. 2019 , eprint=

work page 2019
[36]

Unlocking State-Tracking in Linear

Riccardo Grazzi and Julien Siems and Arber Zela and J. Unlocking State-Tracking in Linear. International Conference on Learning Representations , year=

work page
[37]

2026 , eprint=

A Fully Tunable Ultra-Low Power Current-Mode Memory Cell in Standard CMOS Technology , author=. 2026 , eprint=

work page 2026
[38]

2026 , eprint=

Parallelizable memory recurrent units , author=. 2026 , eprint=

work page 2026
[39]

2026 , note=

Hardware-Software Co-Design of Scalable, Energy-Efficient Analog Recurrent Computations , author=. 2026 , note=

work page 2026
[40]

2026 , note=

On the Importance of Multistability for Horizon Generalization in Reinforcement Learning , author=. 2026 , note=

work page 2026
[41]

Advances in Neural Information Processing Systems , volume=

Combining Recurrent, Convolutional, and Continuous-time Models with Linear State-Space Layers , author=. Advances in Neural Information Processing Systems , volume=

work page
[42]

2018 , eprint=

Parallelizing Linear Recurrent Neural Nets Over Sequence Length , author=. 2018 , eprint=

work page 2018
[43]

, title=

Blelloch, Guy E. , title=. IEEE Transactions on Computers , volume=. 1989 , month=

work page 1989
[44]

Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , pages=

Simple Recurrent Units for Highly Parallelizable Recurrence , author=. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , pages=. 2018 , publisher=

work page 2018
[45]

Maximilian Beck and Korbinian P. x. Advances in Neural Information Processing Systems , volume=

work page
[46]

Advances in Neural Information Processing Systems , volume=

Hierarchically Gated Recurrent Neural Network for Sequence Modeling , author=. Advances in Neural Information Processing Systems , volume=

work page
[47]

Scalable matmul-free language modeling,

Rui-Jie Zhu and Yu Zhang and Steven Abreu and Ethan Sifferman and Tyler Sheaves and Yiqiao Wang and Dustin Richmond and Sumit Bam Shrestha and Peng Zhou and Jason K. Eshraghian , year=. Scalable. 2406.02528 , archivePrefix=

work page arXiv
[48]

Proceedings of the IEEE , volume=

Gradient-based learning applied to document recognition , author=. Proceedings of the IEEE , volume=. 1998 , publisher=

work page 1998
[49]

Krizhevsky, Alex and Hinton, Geoffrey , title=

work page
[50]

and Daly, Raymond E

Maas, Andrew L. and Daly, Raymond E. and Pham, Peter T. and Huang, Dan and Ng, Andrew Y. and Potts, Christopher , title=. Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies , month=. 2011 , address=

work page 2011
[51]

Bowman , booktitle=

Nikita Nangia and Samuel R. Bowman , booktitle=. 2018 , publisher=

work page 2018
[52]

librosa/librosa: 0.11.0 , month=

McFee, Brian and Matt McVicar and Daniel Faronbi and Iran Roman and Matan Gover and Stefan Balke and Scott Seyfarth and Ayoub Malek and Colin Raffel and Vincent Lostanlen and Benjamin van Niekirk and Dana Lee and Frank Cwitkowitz and Frank Zalkow and Oriol Nieto and Dan Ellis and Jack Mason and Kyungyun Lee and Bea Steers and Emily Halvachs and Carl Thom....

work page doi:10.5281/zenodo.15006942 2025
[53]

2018 , eprint=

Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition , author=. 2018 , eprint=

work page 2018
[54]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

Deep Reinforcement Learning that Matters , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

work page
[55]

Proceedings of Machine Learning and Systems , volume=

Accounting for Variance in Machine Learning Benchmarks , author=. Proceedings of Machine Learning and Systems , volume=

work page
[56]

Sensors , volume=

Tiny Machine Learning and On-Device Inference: A Survey of Applications, Challenges, and Future Directions , author=. Sensors , volume=. 2025 , doi=

work page 2025
[57]

Banbury, Colby and Reddi, Vijay Janapa and Torelli, Peter and Holleman, Jeremy and Jeffries, Nat and Kiraly, Csaba and Montino, Pietro and Kanter, David and Ahmed, Sebastian and Pau, Danilo and others , booktitle=

work page
[58]

Warden, Pete and Situnayake, Daniel , year=

work page
[59]

and Mattina, Matthew and Whatmough, Paul N

Fedorov, Igor and Adams, Ryan P. and Mattina, Matthew and Whatmough, Paul N. , booktitle=. 2019 , address=

work page 2019
[60]

Recurrent Neural Networks Hardware Implementation on

Chang, Andre Xian Ming and Martini, Berin and Culurciello, Eugenio , year=. Recurrent Neural Networks Hardware Implementation on. 1511.05552 , archivePrefix=

work page arXiv
[61]

2505.08599 , archivePrefix=

Sebastian Billaudelle and Laura Kriener and Filippo Moro and Tristan Torchet and Melika Payvand , year=. 2505.08599 , archivePrefix=

work page arXiv
[62]

James Bradbury and Roy Frostig and Peter Hawkins and Matthew James Johnson and Chris Leary and Dougal Maclaurin and George Necula and Adam Paszke and Jake VanderPlas and Skye Wanderman-Milne and Qiao Zhang , title=

work page
[63]

Jonathan Heek and Anselm Levskaya and Avital Oliver and Marvin Ritter and Bertrand Rondepierre and Andreas Steiner and Marc van Zee , title=

work page
[64]

2017 , editor =

McMahan, Brendan and Moore, Eider and Ramage, Daniel and Hampson, Seth and Arcas, Blaise Aguera y , booktitle =. 2017 , editor =

work page 2017