Pith. sign in

REVIEW 3 major objections 5 minor 28 references

STAR restores accuracy in stuck-at-faulted memristive networks by augmenting Equilibrium Propagation retraining with a repair nudge toward pre-fault class-conditional activation targets, recovering up to 59 percentage points on a CIFAR-10 C

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 23:26 UTC pith:SBSKHTM7

load-bearing objection Useful repair method with strong empirical results, but the missing ablation means the headline mechanism—stored pre-fault targets—is not actually isolated from generic hidden-layer nudging. the 3 major comments →

arxiv 2607.15415 v1 pith:SBSKHTM7 submitted 2026-07-16 cs.ET

STAR: Astrocyte-Inspired State-Augmented Repair for Supervised Memristive AI Hardware Systems

Yusuf Ahmed Khan , Zhuangyu Han , Abhronil Sengupta This is my paper
classification cs.ET
keywords memristive crossbarstuck-at faultsEquilibrium Propagationastrocyte-inspired repairstate-augmented repairlocal learningfault toleranceactivation targets
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper proposes STAR, a retraining-based repair for memristive crossbar networks that suffer permanent stuck-at faults. The central idea is to record the healthy network's per-class, per-layer mean activations before faults, then retrain the faulted network with Equilibrium Propagation while adding an extra nudging term that pulls hidden states back toward those stored means. The authors report that this restores large fractions of clean accuracy—e.g., a VGG-5 on CIFAR-10 recovering to 72.65% at a 12% fault rate where unassisted retraining reaches 13.34%—with only about 5% of model size needed to store the targets. The work is motivated by a spiking-network study of astrocytic modulation showing repair acts at the population level on surviving synapses. If right, STAR offers a low-overhead, fault-map-free path to self-repair for local, on-chip learning.

Core claim

The paper's central claim is that permanent stuck-at faults in crossbar-mapped networks can be repaired in situ—without locating or replacing faulty weights—by using the healthy network's own equilibrium statistics as a repair guide. After training an EP network fault-free, the authors store the mean free-phase activation of each intermediate layer for each class. When faults are later injected, they retrain the network with EP and augment the nudged-phase dynamics with a quadratic penalty R_h(s,y) = (1/N_h) Σ_n ½‖s^n − μ^n_y‖² that drives each hidden layer's state toward the stored class-conditional target, plus an optional output-layer target. Because the repair signal depends only on loca

What carries the argument

The repair nudge: a quadratic, class-conditional penalty R_h(s,y) added to the EP primitive function during the nudged phases, steering hidden-layer equilibrium states toward pre-fault activation targets μ^n_y recorded from the healthy network. Its own strength β_r (and β_r,out for the output layer) makes it an independent control knob alongside the task nudge β. The mechanism preserves EP's local contrastive update—the weight change still reduces to a difference of products of neuron activations at the two nudged steady states—while shifting those steady states so surviving weights collectively reconstruct pre-fault representations. The paper grounds this design in a computational neuroscie

Load-bearing premise

The load-bearing premise is that a single stored mean of the healthy network's hidden activations per class and layer is a sufficient anchor for repair; if those activations are multimodal or the faults pull the network too far from the anchor, the nudge could distort the representation instead of restoring it.

What would settle it

Train a network with deliberately bimodal within-class hidden activations (e.g., two well-separated clusters per class), record only the class-conditional mean as STAR does, inject stuck-at faults, and run repair. If recovery using the single mean is no better than unassisted retraining—or worse than repair using per-sample or mixture targets—the claim that a mean activation anchor restores pre-fault representations is falsified.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Stuck-at-faulted crossbar networks can be repaired with a stored table of per-class activation means, avoiding fault localization and hardware redundancy for the tested fault rates.
  • The repair nudge keeps the local, contrastive Hebbian structure of EP intact, so the approach stays compatible with in-memory computing constraints.
  • STAR stabilizes retraining under fault stochasticity: reported across-seed variability is much lower than unassisted retraining, which sometimes collapses entirely.
  • The method extends astrocyte-inspired self-repair from small unsupervised spiking benchmarks to supervised deep CNNs and CIFAR-10.
  • Repair-target storage is about 5% of model size (compression ratios 20×–191× across the tested architectures).

Where Pith is reading between the lines

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

  • The single-mean anchor may underperform on classes with multimodal hidden-activation distributions; storing Gaussian mixtures or per-class exemplars and comparing recovery would test this boundary.
  • The same state-augmented nudging idea could apply to other gradual degradation modes—conductance drift, noise, or partial stuck-at—not just the two abrupt stuck-at classes modeled here.
  • At extreme fault rates the optimal repair strength shifts lower (the paper observes this at p_fault=0.9), suggesting an adaptive β_r schedule could push recovery further; this is a direct, testable extension.
  • Because the targets are recorded only from the healthy network, the method behaves like a one-shot distillation of internal representations; one could also refresh targets after partial repair and iterate, which the paper does not explore.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes STAR, a retraining-based repair mechanism for memristive crossbar networks trained with Equilibrium Propagation (EP). It first reports a small computational neuroscience simulation of astrocytic modulation under permanent stuck-at faults, concluding that repair acts at the population level on surviving synapses. STAR then stores per-class, per-layer mean free-phase activations of the healthy network and, during EP retraining on the faulted network, adds a quadratic repair nudge that pulls hidden and output states toward these stored targets. Experiments on MLPs (MNIST) and a VGG-5-like CNN (CIFAR-10) report large accuracy recoveries over unassisted EP retraining, with low variance and modest storage overhead.

Significance. If the proposed mechanism is the actual cause of the reported recoveries, STAR is a practical, low-overhead fault-recovery method for local-learning hardware, and the paper provides a useful evaluation protocol: multiple seeds, detailed hyperparameter tables, a storage-overhead analysis, and a clear hardware fault model. The main limitation is that the paper does not isolate whether the stored pre-fault class-conditional activations are the active ingredient, or whether any hidden-layer nudging term would yield similar gains. The results are consistent and reproducible in design, but the central causal claim is underdetermined by the current experiments.

major comments (3)
  1. [Section V-B and Eqs. (13)-(16)] The claim that the no-STAR baseline 'isolates the contribution of the STAR repair signal' is not justified. Setting beta_r=beta_r,out=0 removes both the stored class-conditional anchors and any hidden-layer nudging. Because Ref. [24] (self-cited) already shows that intermediate error signals improve deep EP training, the large gains in Table I could stem from the mere presence of a hidden-layer nudge rather than from the pre-fault mu_y anchors. Please add ablations: (a) EP with a hidden-layer nudge toward a random/static/global-mean target, (b) [24]-style intermediate error signals without stored means, (c) per-sample activation targets, and (d) a hyperparameter-matched generic quadratic nudge. Without such controls, the paper establishes only that EP with a hidden-layer MSE nudge beats EP alone.
  2. [Section V-B and Table I] Repair strengths beta_r and beta_r,out are reported as 'tuned independently for each architecture and dataset to maximize recovery,' i.e., tuned on the test set; no validation split or selection criterion is described. This risks inflating the reported gains. Please specify a validation-based tuning protocol or, at minimum, report results for a fixed default and for a hold-out validation set. Fig. 4 further shows that beta_r=4 is not optimal at p_fault=0.9, so the Table I numbers do not uniformly reflect the tuned values; this should be clarified.
  3. [Section IV-E, Eq. (17)] The contraction condition is stated as sufficient but the Lipschitz constants K1-K4 are never computed or bounded for the architectures used. Since R_h and R_out are quadratic, K3 and K4 depend on the state norm and are not obviously small; no evidence is given that Eq. (17) holds in the experiments. Empirical convergence is demonstrated, but the formal claim should be qualified as a sufficient condition not verified in practice, or the constants should be estimated.
minor comments (5)
  1. [After Section VII, before References] The passage consisting of '/uni00000015/...' appears to be corrupted reference text and should be removed or reformatted.
  2. [Section III-B] The spiking-network study reports single-trace dynamics in Fig. 3 without error bars or multiple-seed statistics; a brief statement on variability would strengthen the qualitative claim.
  3. [Section VI-B and Table I] Fig. 4 shows that at p_fault=0.9, beta_r=0.5-1 outperforms the beta_r=4 used in Table I. The paper should either adopt fault-rate-aware tuning or note that Table I understates STAR's potential at extreme fault rates.
  4. [Section I, Contributions] The phrase 'MNIST/FMNIST' should be 'MNIST/F-MNIST' for consistency with the later notation.
  5. [Section IV-C] The layer-wise w_max determination is described as a 'percentile analysis' but no percentile value is given; please specify the procedure.

Circularity Check

0 steps flagged

No significant circularity: STAR is judged against external task accuracy; the stored activation targets are not a fitted stand-in for the reported metric. Some same-group citations appear but are not load-bearing.

full rationale

The central claim is that EP retraining augmented by class-conditional activation targets mu_y recovers task accuracy after stuck-at faults. The targets are recorded from the healthy network's free-phase equilibria (Section IV-D, Eq. 13) and used as an additional nudge inside Phi_star (Eq. 14). The reported quantity is held-out classification accuracy (Table I), not closeness to mu_y. Since a retrained network could satisfy the repair loss and still fail to generalize, the recovery gains are not forced by construction; this is an external benchmark, so the derivation is self-contained in the relevant sense. The main caveat is experimental, not circular: Section V-B claims the no-STAR baseline 'isolates the contribution of the STAR repair signal,' but setting beta_r=beta_r,out=0 removes both the stored-target anchor and the entire hidden-layer nudge, so the table does not prove the pre-fault statistics are the active ingredient (a random/static hidden-target ablation is missing). That is a confound/underdetermination, not a reduction of the prediction to the input. Self-citations [16], [17], [18], and [24] appear, but they are background or secondary: [24] is invoked in Section IV-E for a fixed-point argument, which is not the empirical headline and is an extension of a standard EP contraction argument rather than a circular import of the result. No uniqueness theorem is imported, no ansatz is smuggled via citation, and no fitted parameter is relabeled as a prediction. Accordingly, the circularity burden is minimal.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The method's empirical success depends on the augmented EP primitive (borrowed from [24]), the stored pre-fault class-conditional means, and the specific hardware fault model. The neuroscience study is motivational; the algorithm does not use its parameters. Main free parameters are the two repair strengths (tuned on the evaluation data), the conductance clamp settings (wmax percentile, on-off ratio), and the stored activation targets themselves.

free parameters (5)
  • βr (hidden-layer repair nudge strength) = 4 (MLP/MNIST), 1 (CNN/CIFAR-10)
    Tuned per architecture and dataset 'to maximize recovery'; no validation split reported, so this is fitted to the evaluation data.
  • βr,out (output-layer repair nudge strength) = 4 (MLP/MNIST), 0 (CNN/CIFAR-10)
    Tuned separately from βr; set to 0 for CNN, meaning no output-repair term there.
  • wmax (layer-wise maximum conductance) = derived from unspecified layer-wise percentile of trained weights
    Determines SA-2×wmax fault severity and the healthy weight clamp [wmin, wmax]; the percentile choice is not stated.
  • on-off ratio wmax/wmin = 100
    Hand-set to derive wmin; affects the healthy weight clamp range and thus fault impact.
  • Class-conditional activation targets µ^n_y = per-class mean free-phase states per layer
    Recorded from the healthy network's training set; the repair nudge is anchored to these means, so the method's success depends on them.
axioms (5)
  • standard math EP dynamics converge to a fixed point and the contrastive update estimates the gradient when the primitive is smooth (Scellier-Bengio and 3-phase estimator)
    Invoked in Sec IV-A/E; STAR's convergence and gradient claims rest on this, with the augmented quadratic repair terms claimed to preserve Lipschitz continuity.
  • domain assumption The Li-Rinzel astrocyte model and DSE/eSP pathway (Eqs. 1-6) capture astrocytic self-repair principles that transfer to supervised EP networks
    Section III uses a 5-3-2 spiking network to infer a population-level recovery principle; the transfer to EP is asserted, not proven.
  • domain assumption Permanent faults are adequately modeled as uniformly random SA-0 and SA-2×wmax on individual W+/W- branches, equally likely, with healthy weights clamped to [wmin,wmax]
    Section IV-C; real device faults may be correlated, partial, or involve peripheral circuitry, which is not modeled.
  • ad hoc to paper A single per-class mean of the healthy network's free-phase states is a sufficient repair anchor for each layer
    Eq. (13) uses µ^n_y only; within-class multimodal activation structure or per-sample variation is ignored.
  • standard math The contraction condition K1+|β|K2+|βr|K3+|βr,out|K4 < 1 holds for the tested networks, guaranteeing a fixed point
    Eq. (17) in Sec IV-E; the Lipschitz constants are not computed or verified, and the proof is an extension of [24] stated without derivation.

pith-pipeline@v1.3.0-alltime-deepseek · 3639 in / 4203 out tokens · 213072 ms · 2026-08-01T23:26:47.032256+00:00 · methodology

0 comments
read the original abstract

Memristive crossbar arrays have emerged as a promising platform for efficient on-chip learning, enabling local learning rules such as Equilibrium Propagation (EP) to be realized without the memory overhead of conventional backpropagation. However, as these devices age, permanent stuck-at (SA) faults accumulate at individual synaptic elements, irreversibly corrupting the stored weights and degrading model performance in ways that in-situ retraining alone cannot address. Existing mitigation strategies either rely on hardware redundancy at significant area and power cost, or require explicit fault localization that is impractical to perform continuously on-chip. This paper adopts a brain-inspired fault recovery route motivated by self-repair functionalities enabled by astrocytes -- a type of glial cell. We couple a computational neuroscience study of astrocytic neuromodulation under permanent synaptic faults with an algorithmic recovery mechanism for faulted crossbar-mapped networks. Our computational neuroscience study characterizes how astrocytes modulate surviving synapses in a bidirectional EP-specific network setting under both SA-0 and high-conductance stuck-at fault conditions, revealing that recovery operates at the neural population level. Motivated by this observation, we propose a retraining-based repair mechanism that augments EP with an additional repair nudge anchored to pre-fault activation targets of the healthy network, encouraging surviving weights to collectively reconstruct pre-fault internal representations without any information of which weights are faulty. To our knowledge, STAR is the first astrocyte-inspired repair mechanism to operate under a supervised local learning rule and to scale to convolutional networks.

Figures

Figures reproduced from arXiv: 2607.15415 by Abhronil Sengupta, Yusuf Ahmed Khan, Zhuangyu Han.

Figure 1
Figure 1. Figure 1: Overview of the STAR framework. (Left) A crossbar array sub [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Astrocyte-mediated synaptic modulation under synaptic faults. Neu [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Bidirectional synaptic weight dynamics of the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Test accuracy of MLP-1H as a function of repair strength [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Epoch-wise test accuracy across retraining epochs for the VGG-5 CNN on CIFAR-10, expanding upon the converged results reported in Table I across [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 13 canonical work pages

  1. [1]

    Backpropagation and the brain,

    T. P. Lillicrap, A. Santoro, L. Marris, C. J. Akerman, and G. Hinton, “Backpropagation and the brain,”Nature Reviews Neuroscience, vol. 21, no. 6, pp. 335–346, 2020, doi: 10.1038/s41583-020-0277-3

    doi:10.1038/s41583-020-0277-3 2020

  2. [2]

    Theories of error back-propagation in the brain,

    J. C. R. Whittington and R. Bogacz, “Theories of error back-propagation in the brain,”Trends in Cognitive Sciences, vol. 23, no. 3, pp. 235–250, Mar. 2019, doi: 10.1016/j.tics.2018.12.005

    doi:10.1016/j.tics.2018.12.005 2019

  3. [3]

    The recent excitement about neural networks,

    F. Crick, “The recent excitement about neural networks,”Nature, vol. 337, no. 6203, pp. 129–132, Jan. 1989, doi: 10.1038/337129a0

    work page doi:10.1038/337129a0 1989

  4. [4]

    Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain,

    F. A. Azevedo, L. R. Carvalho, L. T. Grinberg, J. M. Farfel, R. E. Ferretti, R. E. Leite, W. Jacob Filho, R. Lent, and S. Herculano-Houzel, “Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain,”J. Comparative Neurol- ogy, vol. 513, no. 5, pp. 532–541, Apr. 2009, doi: 10.1002/cne.21974

    work page doi:10.1002/cne.21974 2009

  5. [5]

    The glia/neuron ratio: how it varies uniformly across brain structures and species and what that means for brain physiology and evolution,

    S. Herculano-Houzel, “The glia/neuron ratio: how it varies uniformly across brain structures and species and what that means for brain physiology and evolution,”Glia, vol. 62, no. 9, pp. 1377–1391, Sep. 2014, doi: 10.1002/glia.22683

    work page doi:10.1002/glia.22683 2014

  6. [6]

    Tripartite synapses: glia, the unacknowledged partner,

    A. Araque, V . Parpura, R. P. Sanzgiri, and P. G. Haydon, “Tripartite synapses: glia, the unacknowledged partner,”Trends in Neurosciences, vol. 22, no. 5, pp. 208–215, 1999, doi: 10.1016/S0166-2236(98)01349-6

    work page doi:10.1016/s0166-2236(98)01349-6 1999

  7. [7]

    Astrocytes gate Hebbian synaptic plas- ticity in the striatum,

    S. Valtcheva and L. Venance, “Astrocytes gate Hebbian synaptic plas- ticity in the striatum,”Nature Communications, vol. 7, p. 13845, Dec. 2016, doi: 10.1038/ncomms13845

    work page doi:10.1038/ncomms13845 2016

  8. [8]

    Integrated brain circuits: astrocytic networks modulate neuronal activity and behavior,

    M. M. Halassa and P. G. Haydon, “Integrated brain circuits: astrocytic networks modulate neuronal activity and behavior,”Annual Review of Physiology, vol. 72, pp. 335–355, 2010, doi: 10.1146/annurev-physiol- 021909-135843

    work page doi:10.1146/annurev-physiol- 2010

  9. [9]

    Self- repair in a bidirectionally coupled astrocyte-neuron (AN) system based on retrograde signaling,

    J. Wade, L. J. McDaid, J. Harkin, V . Crunelli, and S. Kelso, “Self- repair in a bidirectionally coupled astrocyte-neuron (AN) system based on retrograde signaling,”Frontiers in Computational Neuroscience, vol. 6, 2012, doi: 10.3389/fncom.2012.00076

    Pith/arXiv arXiv 2012

  10. [10]

    Endocannabinoid signaling and synaptic function,

    P. E. Castillo, T. J. Younts, A. E. Ch ´avez, and Y . Hashimotodani, “Endocannabinoid signaling and synaptic function,”Neuron, vol. 76, no. 1, pp. 70–81, Oct. 2012, doi: 10.1016/j.neuron.2012.09.020

    work page doi:10.1016/j.neuron.2012.09.020 2012

  11. [11]

    Properties of synaptically evoked astrocyte calcium signal reveal synaptic information processing by astrocytes,

    G. Perea and A. Araque, “Properties of synaptically evoked astrocyte calcium signal reveal synaptic information processing by astrocytes,” Journal of Neuroscience, vol. 25, no. 9, pp. 2192–2203, Mar. 2005, doi: 10.1523/JNEUROSCI.3965-04.2005

    work page doi:10.1523/jneurosci.3965-04.2005 2005

  12. [12]

    Astrocyte calcium signaling: the third wave,

    N. Bazargani and D. Attwell, “Astrocyte calcium signaling: the third wave,”Nature Neuroscience, vol. 19, no. 2, pp. 182–189, Feb. 2016, doi: 10.1038/nn.4201

    work page doi:10.1038/nn.4201 2016

  13. [13]

    Layer ensemble averaging for fault tolerance in memristive neural networks,

    O. Yousufet al., “Layer ensemble averaging for fault tolerance in memristive neural networks,”Nature Communications, vol. 16, no. 1, p. 1250, Feb. 2025, doi: 10.1038/s41467-025-56319-6

    work page doi:10.1038/s41467-025-56319-6 2025

  14. [14]

    Reliability-driven memristive crossbar design in neuro- morphic computing systems,

    Q. Xuet al., “Reliability-driven memristive crossbar design in neuro- morphic computing systems,”IEEE Trans. Autom. Sci. Eng., vol. 20, no. 1, pp. 74–87, Jan. 2023, doi: 10.1109/TASE.2021.3125065

    arXiv 2023

  15. [15]

    Towards biology-inspired fault tolerance of neuromorphic hardware for space applications,

    S. Matinizadeh, S. Johari, A. Mohammadhassani, and A. Das, “Towards biology-inspired fault tolerance of neuromorphic hardware for space applications,” inProc. IEEE Int. Symp. Defect and Fault Tolerance in VLSI and Nanotechnology Systems (DFT), Oct. 2024, pp. 1–7, doi: 10.1109/DFT63277.2024.10753531

    arXiv 2024

  16. [16]

    StochEP: Stochastic equilibrium propagation for spiking convergent recurrent neural networks,

    J. Lin, Y . Jiang, and A. Sengupta, “StochEP: Stochastic equilibrium propagation for spiking convergent recurrent neural networks,” arXiv preprint arXiv:2511.11320, 2025, doi: 10.48550/arXiv.2511.11320

    work page doi:10.48550/arxiv.2511.11320 2025

  17. [17]

    On the self-repair role of astrocytes in STDP enabled unsupervised SNNs,

    M. Rastogi, S. Lu, N. Islam, and A. Sengupta, “On the self-repair role of astrocytes in STDP enabled unsupervised SNNs,”Frontiers in Neuroscience, vol. 14, Jan. 2021, doi: 10.3389/fnins.2020.603796

    arXiv 2021

  18. [18]

    Astromorphic self- repair of neuromorphic hardware systems,

    Z. Han, A. N. M. N. Islam, and A. Sengupta, “Astromorphic self- repair of neuromorphic hardware systems,” inProc. AAAI Conf. Ar- tificial Intelligence, vol. 37, no. 6, pp. 7821–7829, Jun. 2023, doi: 10.1609/aaai.v37i6.25947

    work page doi:10.1609/aaai.v37i6.25947 2023

  19. [19]

    Exploring self-repair in a coupled spiking astrocyte neural network,

    J. Liuet al., “Exploring self-repair in a coupled spiking astrocyte neural network,”IEEE Trans. Neural Netw. Learn. Syst., vol. 30, no. 3, pp. 865–875, Mar. 2019, doi: 10.1109/TNNLS.2018.2854291

    arXiv 2019

  20. [20]

    Adenosine triphosphate-mediated self-repairing astrocyte–neuron network model,

    J. Liu, X. Dong, S. Qiuet al., “Adenosine triphosphate-mediated self-repairing astrocyte–neuron network model,”European Physi- cal Journal Special Topics, vol. 234, pp. 8301–8318, 2026, doi: 10.1140/epjs/s11734-025-01729-2

    work page doi:10.1140/epjs/s11734-025-01729-2 2026

  21. [21]

    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, May 2017, doi: 10.3389/fn- com.2017.00024

    arXiv 2017

  22. [22]

    Scaling equilibrium propagation to deep ConvNets by dras- tically reducing its gradient estimator bias,

    A. Laborieux, M. Ernoult, B. Scellier, Y . Bengio, J. Grollier, and D. Querlioz, “Scaling equilibrium propagation to deep ConvNets by dras- tically reducing its gradient estimator bias,”Frontiers in Neuroscience, vol. 15, 2021, doi: 10.3389/fnins.2021.633674

    arXiv 2021

  23. [23]

    Updates of equilibrium propagation match gradients of backprop through time in an RNN with static input,

    M. Ernoult, J. Grollier, D. Querlioz, Y . Bengio, and B. Scellier, “Updates of equilibrium propagation match gradients of backprop through time in an RNN with static input,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 32, 2019

    2019

  24. [24]

    Scalable equilibrium propagation via intermediate error signals for deep convolutional CRNNs,

    J. Lin, M. Bal, and A. Sengupta, “Scalable equilibrium propagation via intermediate error signals for deep convolutional CRNNs,” arXiv preprint arXiv:2508.15989, 2025

    Pith/arXiv arXiv 2025

  25. [25]

    On improving fault tolerance of memristor crossbar based neural network designs by target sparsifying,

    S. Jin, S. Pei, and Y . Wang, “On improving fault tolerance of memristor crossbar based neural network designs by target sparsifying,” inProc. Design, Automation & Test in Europe Conf. (DATE), Grenoble, France, Mar. 2020, pp. 91–96, doi: 10.23919/DATE48585.2020.9116187

    arXiv 2020

  26. [26]

    Handling stuck-at-faults in memristor crossbar arrays using matrix transformations,

    B. Zhang, N. Uysal, D. Fan, and R. Ewetz, “Handling stuck-at-faults in memristor crossbar arrays using matrix transformations,” inProc. 24th Asia South Pacific Design Automation Conf. (ASP-DAC), Tokyo, Japan, Jan. 2019, pp. 438–443, doi: 10.1145/3287624.3287707

    arXiv 2019

  27. [27]

    Zero-overhead protection for CNN weights,

    S. Burel, A. Evans, and L. Anghel, “Zero-overhead protection for CNN weights,” inProc. IEEE Int. Symp. Defect and Fault Tolerance in VLSI and Nanotechnology Systems (DFT), Athens, Greece, Oct. 2021, pp. 1–6, doi: 10.1109/DFT52944.2021.9568363

    arXiv 2021

  28. [28]

    Late breaking result: FPGA-based emulation and fault injection for CNN inference accelerators,

    M. H. Ahmadilivaniet al., “Late breaking result: FPGA-based emulation and fault injection for CNN inference accelerators,” arXiv preprint arXiv:2501.12818, Jan. 2025

    Pith/arXiv arXiv 2025