Preprocessing Methods for Memristive Reservoir Computing for Image Recognition

David Saad; Duna Wattad; Rishona Daniels; Ronny Ronen; Shahar Kvatinsky

REVIEW 3 major objections 5 minor 23 references

Preprocessing Methods for Memristive Reservoir Computing for Image Recognition

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read XORing adjacent image rows lifts memristive reservoir accuracy by up to 6%.

desk verdict A useful, reproducible comparison of memristive RC preprocessing methods with a simple parity trick, but the accuracy gain is confounded by reservoir size and lacks error bars. read the letter →

arxiv 2506.05588 v3 pith:5RX4JXPG submitted 2025-06-05 cs.NE cs.ARcs.ET

Rishona Daniels , Duna Wattad , Ronny Ronen , David Saad , Shahar Kvatinsky This is my paper

classification cs.NEcs.ARcs.ET

keywords reservoircomputingmemristorpreprocessingparityMNISTimagerecognitiondelayfeedbacknetworkneuromorphic

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

The paper tries to show that input preprocessing matters as much as reservoir size for memristive reservoir computing, and proposes a parity-based preprocessing method. The method XORs adjacent rows of a binary image and feeds the resulting outline rows into extra volatile-memristor reservoir nodes alongside the original rows. On MNIST, this lifts test accuracy by roughly 6% in one-dimensional input mode and 2% in two-dimensional mode, with only a modest increase in device count. A systematic comparison also shows that sectioning input rows into smaller segments is essential for accuracy, while parity adds a further edge. The practical stake is that cheap preprocessing can substitute for a larger physical reservoir.

What carries the argument

The central object is parity preprocessing: an XOR operation between adjacent rows of the binarized image that produces an extra set of parity rows appended below the image. These rows are converted into write-voltage pulse trains and applied to separate volatile memristors in the reservoir, in addition to the original rows and, in 2D mode, columns. The mechanism works because XOR of two adjacent rows responds only where the rows differ, yielding a sparse outline of the digit that highlights edges; these parity rows contain fewer ones than the original data, so the added energy cost is sublinear. The reservoir itself is a delay feedback network of volatile memristors, whose internal state variable w integrates input pulses according to a fitted metal-oxide model and decays with a time constant tau, giving the echo-state and fading-memory properties needed for temporal pattern projection.

What would settle it

Apply the same parity preprocessing to a dataset whose rows are permuted or statistically independent, or shuffle each MNIST row order independently for every image, and measure test accuracy; if accuracy no longer improves by 2–6%, the row-correlation assumption is confirmed as the mechanism. A concrete calculation is to run the 1D+parity configuration on MNIST with rows randomly reordered within each image and compare to the unshuffled result.

Watch

Extended reading notes

Core claim

The central claim is that parity preprocessing — an XOR of the i-th and (i+1)-th rows of the binarized input image, appended below the original image and converted into separate pulse trains — improves the classification accuracy of a delay-feedback memristive reservoir on MNIST by approximately 6% in the 1D configuration and 2% in the 2D configuration over the same configuration without parity. The parity rows provide a sparse outline of the digit, making edges easier for the reservoir to separate, and they require fewer write pulses because parity rows contain fewer ones than the original data. Compared with the prior delay-feedback memristive system used as the baseline, the parity method reaches 1–2% higher accuracy with fewer devices; compared with a two-memristor-one-capacitor kernel approach, accuracy is slightly lower but without the area and scalability overhead of large capacitors.

Load-bearing premise

The load-bearing premise is that adjacent rows of the input image are statistically similar enough that their XOR forms a useful sparse outline of the shape; for datasets without such row-to-row structure, the parity gain should shrink or vanish.

Editorial extensions

If this is right

Accuracy gains of 2–6% are achievable on MNIST from preprocessing alone, without training more readout weights or enlarging the reservoir beyond the extra parity memristors.
Because parity rows are sparser than the original image, the energy increase from adding parity is sublinear in the added device count, so the accuracy gain does not come at a proportional energy cost.
Sectioning remains the dominant accuracy factor: without it, accuracy stays below 75%, and parity adds its gain on top of sectioning.
The comparison suggests that for the same accuracy, a memristive system with parity preprocessing can use fewer devices than the baseline delay-feedback design, which supports scalability of physical reservoir computing systems.

Reading between the lines

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

If the governing mechanism is that adjacent-row correlation produces useful outlines, the same gain should appear on other digit or character datasets with vertical structure, and should shrink on datasets with rows shuffled or statistically independent rows.
The parity idea could be extended beyond rows: XORing adjacent columns, or combining row and column parity, would add diagonal-edge information that this paper does not test; the paper itself flags cross-row and column operations as future work.
Because quantization to 4 bits preserved accuracy, parity preprocessing may combine well with aggressive analog-to-digital conversion in hardware, an interaction the paper does not explicitly study.
One could try parity at multiple scales — for example, XORing rows at strides of 2, 4, or 8 — to see whether thicker or thinner outlines are better; the paper only tests adjacent-row parity.

Share X Bluesky LinkedIn Reddit HN

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

The paper is a clean, systematic comparison of preprocessing methods for memristive delay-feedback reservoir computing on MNIST, and it adds one genuinely new preprocessing idea: XOR adjacent rows to create 'parity rows' that are fed to additional reservoir memristors. The authors report 2–6% accuracy gains from parity, plus trade-off plots for throughput, energy, and area that hardware people will find useful.

What I like: the simulation setup is described in enough detail to reproduce (equations, parameters, quantization, 500 epochs), the comparison across 1D/2D/sectioning/parity is organized, and the paper is honest that parity is a first step—it explicitly says more sophisticated cross-row/column operations could be explored. The energy result—parity rows contain fewer ones, so energy grows sub-linearly with added devices—is a nice practical observation.

The soft spot is real and it is the one the stress-test flagged. Parity adds (n−1) rows, so the reservoir gets larger. For 1D with four sections, the device count goes from 112 to 220. The paper itself shows that sectioning alone improves accuracy, so the 6% gain from parity is at least partially a capacity/load effect, not necessarily an edge-detection effect. There is no control that holds the number of memristors fixed and swaps parity rows for another equally sized feature set (random sparse rows, duplicated rows, extra sections). The sentence in Section IV claiming the parity rows 'provide a sparse outline' is an interpretation, not a demonstrated mechanism. That should be flagged but not exaggerated: the method still works, whatever the reason.

Also, no multiple runs or error bars anywhere. A 2–6% difference on MNIST could shift with seed or initialization, and the comparison to Du et al. [10] is not controlled (different code, different settings). These are standard empirical-hygiene issues, not fatal ones.

Bottom line: this is a decent engineering study for a niche community. The central claim about parity accuracy gain is plausible but under-supported. A serious referee should ask for an equal-capacity control and repeated runs. I would not desk-reject it; with those additions it could be a solid contribution to the memristive RC literature.

Referee Report

3 major / 5 minor

Summary. This paper evaluates preprocessing methods for delay-feedback memristive reservoir computing on the MNIST dataset. It compares 1D row-wise encoding, 2D row-plus-column encoding, input sectioning, and a proposed parity method that appends XORed adjacent rows to the input. Using a deterministic metal-oxide volatile memristor model, the authors report that parity improves test accuracy by about 6% for 1D and 2% for 2D relative to the same methods without parity, with a modest increase in device count, and they analyze the trade-offs among accuracy, throughput, energy, and area.

Significance. The systematic comparison of preprocessing methods and the proposed parity preprocessing are potentially useful for memristive reservoir computing design. The paper clearly specifies the device model, equations, and parameter choices, and it makes a concrete falsifiable empirical claim about accuracy gains. However, the central claim is currently weakened by a confound between preprocessing content and reservoir size and by the absence of statistical error analysis. If the confound is resolved with suitable control experiments, the result would be a solid empirical contribution; in its present form, the causal attribution to edge information is not established.

major comments (3)

[V-A, Table I] The central claim that parity preprocessing improves accuracy by highlighting image edges is confounded with increased reservoir size. Under 1D preprocessing with k sections, the reservoir has n×k memristors, whereas 1D+parity has (2n−1)×k memristors; for MNIST with k=4 this is 112 versus 220 memristors, and the paper itself states in Section V-A that accuracy rises with sectioning, that is, with more memristors. No experiment in Fig. 4 holds the number of memristors fixed while replacing the XOR rows with equally sized non-edge content such as random sparse rows or duplicated rows. Without such a control, the gain attributed to the 'sparse outline' interpretation in Section IV is not established.
[V-A] All accuracy results in Fig. 4 and the stated 2–6% gains are reported as single runs without error bars, confidence intervals, or random seed information. Because the readout weights are randomly initialized and trained using SGD (Section V), the reported differences may be within run-to-run variability. The authors should report means and standard deviations over multiple trials and, if possible, a statistical significance test. The comparison to Du et al. [10] in Section V-A should also state whether the same data encoding, training epochs, and sectioning were used for both systems.
[V, Equations (1)–(5)] The statement that 6-bit quantization of read currents 'helps assess the realistic accuracy that can be achieved in the presence of memristor variability' is not supported by the model. Equations (1)–(5) are deterministic, and no device-to-device or cycle-to-cycle variability is introduced; quantization models only ADC resolution. The authors should either incorporate an explicit variability model or rephrase the claim to refer only to readout quantization.

minor comments (5)

[References] The reference list contains visible LaTeX artifacts such as 'V ol.', 'Y .', and 'et al.,,'; these should be cleaned in the final version.
[I] The text mentions 'Lorentz-63'; the standard name is Lorenz-63.
[V-A] The phrase 'over 4 Sections' is unclear; please specify the exact section counts used for the reported average improvements.
[V] The paper would be easier to reproduce if the MATLAB implementation details, including readout initialization and random seed procedure, were provided or if the code were released.
[Fig. 5] The green and blue boxes in Fig. 5 are helpful but not defined in the caption; please explain which configurations are marked and why.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the parity preprocessing gain is an empirical comparison, not a derivation from fitted quantities.

full rationale

The paper is an empirical evaluation, not a derivation chain. The proposed parity preprocessing is a fixed XOR transform applied to adjacent image rows (Section IV); it contains no fitted parameters and is not defined in terms of the resulting test accuracy. The memristive model parameters (alpha, beta, gamma, delta, lambda, eta, tau) are taken from the external model of [18], and the reservoir readout is trained only on the 60,000 training images and evaluated on the separate 10,000-image test set (Section V). The claimed 2-6% gain is a measured difference between preprocessing pipelines, not a quantity that equals an input by construction. Citations to prior work ([10], [18], [23]) are external; no load-bearing result rests on a self-citation, and no uniqueness theorem or ansatz is imported from the authors' own prior work. The concern that the parity gain may be confounded with increased reservoir size (no equal-device-count control in Fig. 4) is a legitimate experimental limitation, but it is not circularity: the reported accuracy is not forced by definition, and the causal claim about edge highlighting is an interpretation, not part of a derivation chain. Therefore the paper is self-contained against the circularity patterns defined here.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The central claim depends on hand-chosen simulation hyperparameters (tau, voltages, threshold, quantization) and on the domain assumptions that the memristor model is accurate, binarization preserves information, and the XOR edge representation is useful. No new physical entities are introduced.

free parameters (8)

tau (diffusion time constant) = 5 ns
Chosen for the volatile memristor; controls the rate of state decay and thus the reservoir's fading memory. The paper states the decay rate 'needs to be properly tuned' but does not describe a tuning procedure.
Vwrite = 1.5 V
Write pulse voltage used to update memristor state; a hand-chosen simulation setting.
Vread = 0.6 V
Read pulse voltage; chosen lower than Vwrite to avoid state change.
tpulse = 1 ns
Pulse width for both write and read operations.
Binarization threshold = 25
Greyscale pixels above 25 are set to '1'; chosen to retain digit shape.
Readout learning rate = 0.02
Hyperparameter for SGD training of the readout layer.
Number of training epochs = 500
Training duration for readout.
Read current quantization bits = 6 bits
The paper analyzes 1-7 bits and uses 6 bits for the main results; quantization choice affects accuracy.

assumptions (4)

domain assumption The memristor model of Chang et al. [18] accurately describes the behavior of the simulated volatile memristor.
The entire simulation uses Equations (1)-(5) from [18]; if the model is inaccurate, all reported accuracies and energy estimates are unreliable.
domain assumption Binarizing MNIST at threshold 25 preserves the shape of the digits well enough for the preprocessing methods to be fairly compared.
The paper binarizes pixels above 25 to 1; this choice affects the pulse trains and thus the reservoir states.
domain assumption The echo-state and fading-memory properties of the memristor reservoir are sufficient to produce useful representations for image classification.
This is the foundation of reservoir computing; the paper relies on it without proof.
ad hoc to paper XOR of adjacent rows yields a sparse outline of the digit, and this representation is useful for classification.
This is the core assumption behind the proposed parity method; the paper states it in Section IV but provides no theoretical or empirical justification beyond the accuracy results.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Preprocessing Methods for Memristive Reservoir Computing for Image Recognition." pith.science (2026). https://pith.science/paper/5RX4JXPG

@misc{pith2026250605588,
  author       = {Pith},
  title        = {Pith review of: Preprocessing Methods for Memristive Reservoir Computing for Image Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5RX4JXPG}},
  note         = {Machine review of arXiv:2506.05588}
}

read the original abstract

Reservoir computing (RC) has attracted attention as an efficient recurrent neural network architecture due to its simplified training, requiring only its last perceptron readout layer to be trained. When implemented with memristors, RC systems benefit from their dynamic properties, which make them ideal for reservoir construction. However, achieving high performance in memristor-based RC remains challenging, as it critically depends on the input preprocessing method and reservoir size. Despite growing interest, a comprehensive evaluation that quantifies the impact of these factors is still lacking. This paper systematically compares various preprocessing methods for memristive RC systems, assessing their effects on accuracy and energy consumption. We also propose a parity-based preprocessing method that improves accuracy by 2-6% while requiring only a modest increase in device count compared to other methods. Our findings highlight the importance of informed preprocessing strategies to improve the efficiency and scalability of memristive RC systems.

Figures

Figures reproduced from arXiv: 2506.05588 by the authors.

**Figure 2.** Schematic of delay feedback network (DFN) reservoir computing [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Schematics of preprocessing methods. (a) One-dimensional preprocessing: each horizontal row of pixels is converted into a pulse train of write voltages [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Test accuracy across various preprocessing methods. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Trade-offs between accuracy and (a) throughput, (b) energy efficiency, and (c) area. Methods using a single section with accuracy below 75% are [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages

[10]

Reservoir computing using dynamic memristors for temporal information processing,

C. Du et al. , “Reservoir computing using dynamic memristors for temporal information processing,” Nat. Commun., V ol. 8, No. 1, p. 2204, Dec. 2017

work page 2017
[1]

Next generation reservoir computing,

D. J. Gauthier, E. Bollt, A. Griffith, and W. A. S. Barbosa, “Next generation reservoir computing,” Nat. Commun., V ol. 12, No. 1, p. 5564, Dec. 2021

work page 2021
[2]

A combination forecasting model of wind speed based on decomposition,

Z. Tian, H. Li, and F. Li, “A combination forecasting model of wind speed based on decomposition,” Energy Rep. , V ol. 7, pp. 1217–1233, Nov. 2021

work page 2021
[3]

Stock market index prediction based on reservoir computing models,

W.-J. Wang, Y . Tang, J. Xiong, and Y .-C. Zhang, “Stock market index prediction based on reservoir computing models,” Expert Syst. Appl. , V ol. 178, pp. 115022, Sept. 2021

work page 2021
[4]

Tong and G

Z. Tong and G. Tanaka, ”Reservoir computing with untrained convolu- tional neural networks for image recognition,” Proc. Int. Conf. Pattern Recognit. (ICPR), pp. 1289-1294, Beijing, China, Aug. 2018

work page 2018
[5]

Information processing using a single dynamical node as complex system,

L. Appeltant et al. , “Information processing using a single dynamical node as complex system,” Nat. Commun., V ol. 2, No. 1, pp. 1–6, Sept. 2011

work page 2011
[6]

Prediction of chaotic time series using recurrent neural networks and reservoir computing techniques: A comparative study,

S. Shahi, F. H. Fenton, and E. M. Cherry, “Prediction of chaotic time series using recurrent neural networks and reservoir computing techniques: A comparative study,” Mach. Learn. Appl. , V ol. 8, pp. 100300, June 2022

work page 2022
[7]

Physical reservoir computing with emerging electron- ics,

X. Liang et al.,, “Physical reservoir computing with emerging electron- ics,” Nat. Electron., V ol. 7, pp. 193-206, Mar. 2024

work page 2024

Show all 23 references

[8]

Delay-based reservoir computing: Noise effects in a combined analog and digital implementation,

M. C. Soriano et al. , “Delay-based reservoir computing: Noise effects in a combined analog and digital implementation,” IEEE Trans. Neural Netw. Learn. Syst. , V ol. 26, No. 2, pp. 388–393, Feb. 2015

2015
[9]

Liquid state machine based pattern recognition on FPGA with firing-activity dependent power gating and approximate computing,

Q. Wang, Y . Li, and P. Li, “Liquid state machine based pattern recognition on FPGA with firing-activity dependent power gating and approximate computing,” Proc. IEEE Int. Symp. Circuits Syst. (ISCAS) , pp. 361-364, Montreal, QC, Canada, May 2016

2016
[11]

Advances in photonic reservoir computing,

G. Van Der Sande, D. Brunner, and M. C. Soriano, “Advances in photonic reservoir computing,” Nanophotonics, V ol. 6, No. 3., pp. 561–576, May 2017

2017
[12]

Passive frustrated nanomagnet reservoir comput- ing,

A. J. Edwards et al., “Passive frustrated nanomagnet reservoir comput- ing,” Commun. Phys., V ol. 6, No. 1, p. 215, Dec. 2023

2023
[13]

Echo state graph neural networks with analogue random resistive memory arrays,

S. Wang et al., “Echo state graph neural networks with analogue random resistive memory arrays,” Nat. Mach. Intell., V ol. 5, No. 2, pp. 104–113, Feb. 2023

2023
[14]

Reconfigurable halide perovskite nanocrystal mem- ristors for neuromorphic computing,

R. A. John et al., “Reconfigurable halide perovskite nanocrystal mem- ristors for neuromorphic computing,” Nat. Commun., V ol. 13, No. 1, p. 2074, Dec. 2022

2022
[15]

Nakajima and I

K. Nakajima and I. Fischer, Natural Computing Series Reservoir Com- puting Theory, Physical Implementations, and Applications, Springer Singapore, Aug. 2021

2021
[16]

H. Jaeger, ”The ”echo state” approach to analysing and training recur- rent neural networks-with an erratum note’,” Bonn, Germany: German National Research Center for Information Technology GMD Technical Report, 148, Jan. 2001

2001
[17]

G. D. Zhou et al. , ”V olatile and nonvolatile memristive devices for neuromorphic computing,” Adv. Electron. Mater. , V ol. 8, No. 7, p. 2101127, Feb. 2022

2022
[18]

Chang, et al

T. Chang, et al. , ”Synaptic behaviors and modeling of a metal oxide memristive device,” Appl. Phys. A , V ol. 102, pp. 857–863, Feb. 2011

2011
[19]

Sun et al

J. Sun et al. , ”A novel memristors based echo state network model inspired by the brain’s uni-hemispheric slow-wave sleep characteristics,” Cogn. Comput., V ol. 16, pp. 1470–1483, June 2024

2024
[20]

A. M. Hassan, H. H. Li and Y . Chen, ”Hardware implementation of echo state networks using memristor double crossbar arrays,” Proc. 2017 Int. Joint Conf. Neural Networks (IJCNN) , pp. 2171-2177, Anchorage, AK, USA, May 2017

2017
[21]

Henderson et al., ”Memristor based liquid state machine with method for in-situ training,” IEEE Trans

A. Henderson et al., ”Memristor based liquid state machine with method for in-situ training,” IEEE Trans. Nanotechnol. , V ol. 23, pp. 376-385, Mar. 2024

2024
[22]

Wu et al

X. Wu et al. , ”Nonmasking-based reservoir computing with a single dynamic memristor for image recognition,” Nonlinear Dyn. , V ol. 12, pp. 6663–6678, Mar. 2024

2024
[23]

2Memristor-1Capacitor integrated temporal kernel for high-dimensional data mapping,

S. K. Shim et al., “2Memristor-1Capacitor integrated temporal kernel for high-dimensional data mapping,” Small, V ol. 20, No. 25, p. 2306585, June 2024

2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [10]

Reservoir computing using dynamic memristors for temporal information processing,

C. Du et al. , “Reservoir computing using dynamic memristors for temporal information processing,” Nat. Commun., V ol. 8, No. 1, p. 2204, Dec. 2017

work page 2017

[2] [1]

Next generation reservoir computing,

D. J. Gauthier, E. Bollt, A. Griffith, and W. A. S. Barbosa, “Next generation reservoir computing,” Nat. Commun., V ol. 12, No. 1, p. 5564, Dec. 2021

work page 2021

[3] [2]

A combination forecasting model of wind speed based on decomposition,

Z. Tian, H. Li, and F. Li, “A combination forecasting model of wind speed based on decomposition,” Energy Rep. , V ol. 7, pp. 1217–1233, Nov. 2021

work page 2021

[4] [3]

Stock market index prediction based on reservoir computing models,

W.-J. Wang, Y . Tang, J. Xiong, and Y .-C. Zhang, “Stock market index prediction based on reservoir computing models,” Expert Syst. Appl. , V ol. 178, pp. 115022, Sept. 2021

work page 2021

[5] [4]

Tong and G

Z. Tong and G. Tanaka, ”Reservoir computing with untrained convolu- tional neural networks for image recognition,” Proc. Int. Conf. Pattern Recognit. (ICPR), pp. 1289-1294, Beijing, China, Aug. 2018

work page 2018

[6] [5]

Information processing using a single dynamical node as complex system,

L. Appeltant et al. , “Information processing using a single dynamical node as complex system,” Nat. Commun., V ol. 2, No. 1, pp. 1–6, Sept. 2011

work page 2011

[7] [6]

Prediction of chaotic time series using recurrent neural networks and reservoir computing techniques: A comparative study,

S. Shahi, F. H. Fenton, and E. M. Cherry, “Prediction of chaotic time series using recurrent neural networks and reservoir computing techniques: A comparative study,” Mach. Learn. Appl. , V ol. 8, pp. 100300, June 2022

work page 2022

[8] [7]

Physical reservoir computing with emerging electron- ics,

X. Liang et al.,, “Physical reservoir computing with emerging electron- ics,” Nat. Electron., V ol. 7, pp. 193-206, Mar. 2024

work page 2024