Pith. sign in

REVIEW 4 major objections 5 minor 27 references

In-Sensor Motion Recognition with Memristive System and Light Sensing Surfaces

T0 review · 4 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 Integrating a photovoltaic light-sensing surface with an HfO2 memristive crossbar and a winner-take-all circuit classifies four hand-gesture directions with 97.22% accuracy under 5% noise, at 0.952 nJ per classification.

desk verdict Clever integration and careful energy accounting, but the 97.22% accuracy claim is a tautology because the test reuses the exact templates used to program the weights. read the letter →

arxiv 2506.06829 v1 pith:LM2XCA7D submitted 2025-06-07 cs.HC

Hritom Das , Imran Fahad , SNB Tushar , Sk Hasibul Alam , Graham Buchanan , Danny Scott , Garrett S. Rose , Sai Swaminathan This is my paper · ORCID
classification cs.HC
keywords in-sensorcomputingmemristivecrossbarlight-sensingsurfacephotovoltaiccellshandgesturerecognitionwinner-take-allcircuitHfO2synapseenergy-efficientedgeAI
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes an end-to-end motion-recognition pipeline that keeps both sensing and classification in the analog domain: a light-sensing surface made of photovoltaic cells produces a voltage waveform for each hand-gesture direction, and a $4\times8$ crossbar of $\mathrm{HfO}_2$-based memristive synapses, ranked by a winner-take-all circuit, turns that waveform into a class label. The authors report that mapping the eight-sample gesture templates into the crossbar costs $4.17$ nJ and each classification costs $0.952$ nJ, with $97.22\%$ accuracy maintained when Gaussian noise perturbs the column currents by up to 5%. The motivation is that such a system could run continuously on harvested light, without an ADC, a wireless link, or a battery, while keeping raw sensor data local.

What carries the argument

The load-bearing mechanism is the computational light-sensing surface: photovoltaic cells connected so their voltages cancel in steady light, producing a nonzero algebraic sum only when a moving shadow falls unevenly across them. This surface behaves like an edge-detection kernel, modeled as an asymmetric $4\times4$ matrix, and performs the weighted linear combination directly in the analog domain at $\mathcal{O}(1)$ complexity rather than $\mathcal{O}(M)$ per photodiode. The second mechanism is the $4\times8$ memristive crossbar of current-controlled $\mathrm{HfO}_2$ synapses, programmed by 4-bit DACs to hold the eight-sample gesture templates as resistance levels; the crossbar acts as a dot-product engine whose column currents accumulate the match between an incoming sample pattern and each stored template. The winner-take-all circuit completes the classification by selecting the column with the maximum current.

What would settle it

Record a new set of hand-gesture trials at the same lighting and speed ranges, sample each into the same eight-point format, run them through the memristive crossbar and winner-take-all circuit, and count how often the correct column wins; if accuracy on these held-out trials falls well below 97.22%, the stated robustness claim is not supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that hand-gesture direction can be recognized end-to-end in the analog domain by pairing a computational light-sensing surface with a memristive processing unit. Four pairs of KXOB25-05X3F-TR photovoltaic cells are wired with half in opposite polarity, so steady light cancels the net voltage and an occluding hand produces a time-resolved sequence such as $[0,1,0,-1]$ for a given motion direction. Eight samples of each of the four waveforms are converted by 4-bit DACs into resistance levels between about $5$ and $96\,\mathrm{k}\Omega$ on the columns of a $4\times8$ $\mathrm{HfO}_2$ synaptic crossbar; during testing, the summed column currents are compared by a winner-take-all circuit, and the column with the largest current names the gesture. The authors report $4.17$ nJ average energy for weight mapping, $0.952$ nJ per classification including sampling and readout, and $97.22\%$ correct classification across 81 noise scenarios per gesture type at up to 5% Gaussian noise on the column currents.

Load-bearing premise

The 97.22% accuracy claim assumes that the test inputs—the same eight-sample gesture patterns used to set the device's weights, with noise added only to the final summed currents—represent how the system will behave on real, previously unseen hand movements.

Editorial extensions

If this is right

  • A battery-free always-on gesture sensor becomes plausible: at $0.952$ nJ per classification plus an occasional $4.17$ nJ weight update, a small solar cell or capacitor could sustain continuous monitoring.
  • Because the gesture waveform is already the feature vector, the system can skip analog-to-digital conversion and wireless transmission, removing two of the largest power consumers in typical IoT motion sensing.
  • Local analog processing means raw optical data never leaves the surface, so the architecture provides privacy by construction for ambient activity sensing.
  • The column-per-template organization means adding a new gesture direction is a matter of programming one more crossbar column and letting the winner-take-all circuit rank it with the others.
  • The paper's comparisons show roughly $92\times$ and $21\times$ lower energy than the two prior neuromorphic classifiers, although the systems differ in sensor type and process technology.

Reading between the lines

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

  • My inference: the noise study perturbs the summed column currents rather than the raw optical waveforms, so real-world variation in hand speed, height above the surface, and ambient light would probably reduce accuracy more than the reported 5% noise scenario.
  • My inference: because the same eight-sample templates that set the weights also serve as the test inputs, the 97.22% figure is best understood as a check of the analog read path's consistency, not as a generalization measurement over new users or trials.
  • My inference: the $\mathcal{O}(1)$ sensing complexity suggests the photovoltaic surface can scale to larger areas cheaply, but the crossbar energy and area will grow roughly linearly with the number of stored gesture templates.
  • My inference: a natural next experiment is to record a held-out set of gestures from several users under varied lighting and speed, map them through the same pre-processing, and measure winner-take-all accuracy; that would test whether the accuracy claim generalizes.
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.

Referee Report

4 major / 5 minor

Summary. The paper proposes an end-to-end motion recognition architecture that couples a computational light-sensing surface with an HfO2-based memristive crossbar and a winner-take-all circuit. The authors report a classification energy of 0.952 nJ per gesture and an accuracy of 97.22% under 5% noise for four hand-gesture directions. The central empirical claims rest on an evaluation protocol in which the same eight-sample waveforms that define the programmed column weights are also used as the test inputs, so the reported accuracy is a self-consistency check rather than a measure of recognition performance on unseen gestures.

Significance. If the accuracy and energy claims were properly validated, the architecture would be a meaningful step toward energy-efficient, privacy-preserving motion recognition at the edge, with a credible comparison to prior neuromorphic systems. The paper's strength is its detailed hardware design and the integration of a computational light-sensing surface with a memristive processing unit. However, the validation methodology is circular: the test set is the training set, and the noise model is a contrived adversarial perturbation rather than a realistic stochastic model. As a result, the headline 97.22% accuracy does not support the paper's claims of classification robustness and generalization. The architecture may still be valuable, but the evidence presented does not establish the stated performance.

major comments (4)
  1. [Section IV and Table III] The test protocol reuses the same eight-sample waveforms from Table I that define the column weights in Table II. When the incoming signal is declared to be, for example, BT, the read activation pattern is exactly the binarization of the BT template, so the BT column must produce the maximum current. Correct classification is therefore guaranteed by construction in the noiseless case, and the reported accuracy is not evidence of generalization to new gesture trials. The authors should evaluate with independently collected gesture samples not used in programming (e.g., multiple trials per participant, cross-validation) and report the resulting confusion matrix.
  2. [Section V] The robustness evaluation does not test random noise. The paper states that Gaussian noise is added to the column currents "in such a way that the maximum column current decreases by 1%, 3%, and 5% due to noise, while the current of other columns increases by 1%, 3%, and 5%". This is a deterministic, adversarial margin perturbation, not random noise. In a physical system, noise would affect the analog sensor waveform before the 0.45 V threshold, the device conductances, and the read currents in a correlated, sign-varying manner. The 97.22% figure, computed from 315 of 324 hand-constructed scenarios, is an artifact of this test. The authors should define a realistic noise source (sensor noise, device variability, or circuit noise) and evaluate classification accuracy over many Monte Carlo draws, or measure the system with real noisy signals.
  3. [Section IV] The read activation thresholds each sampled value at 0.45 V, converting the eight samples into a binary pattern and discarding amplitude information. Combined with the fact that the weights are programmed from the same templates, the system is effectively a binary template matcher. The paper should state this limitation clearly, and either demonstrate that the analog amplitude information contributes to classification beyond the binary threshold or reframe the contribution as a template-matching lookup rather than a learned classifier with generalization claims.
  4. [Abstract and Section III] The mapping energy of 4.17 nJ cited in the abstract is not derived in the main text. Section III reports an average DAC energy of 282.56 pJ per column and an average column mapping energy of 3.89 nJ, but it is not clear how these combine to 4.17 nJ or whether the latter is per column or for the full four-column mapping. The authors should provide a consistent energy breakdown with equations and explicitly state what is included in each reported number.
minor comments (5)
  1. [Throughout] There are several typographical and formatting issues, including "memeristive" in Section I, "F ORM", "RESET", and "SET" with spacing artifacts, and "Specter" which should be "Spectre".
  2. [Section I.B, Eq. (1)] The summation in Equation (1) uses V_n in the term but V_i in the indexing; the notation should be made consistent.
  3. [Section I and Section V] The introduction describes "uniformly random noises" while Section V applies a directional worst-case perturbation; the terminology should be aligned so that the noise model is described consistently.
  4. [Table IV] The energy comparison with prior work should state the process node, sensor type, and what is included in each energy figure, since the compared systems use different sensing modalities and may count different peripherals.
  5. [Section III] The statement that "8 random samples will be selected" is ambiguous; Table I appears to show deterministic, time-aligned samples, so the sampling procedure should be clarified.

Circularity Check

2 steps flagged · score 6.0 of 10

Headline accuracy is a self-referential template-matching check, not an externally benchmarked prediction; the 97.22% robustness figure is generated from an author-defined worst-case noise scenario.

  1. self definitional [Section III (Analog Signal Mapping Approach) and Section IV (Testing Approach), Tables I-III]
    "For each object motion class, data will be sampled off-chip, and 8 random samples will be selected to represent a specific movement. ... Then, 8 DACs are utilized to map the eight synaptic elements of a column with 8 sampled pulses, which are different in amplitudes. ... According to Table III, only the column matched with sensor data will produce the maximum amount of current among all the columns."

    The column weights are the DAC-quantized amplitudes of each gesture's eight samples (Table I to Table II), and the test input is the same gesture signal, thresholded at 0.45 V into the binary READ pattern (Section IV). Therefore the matching column's current is maximized by construction: the crossbar is not learning a decision boundary but performing a look-up of stored templates against the same templates. The noiseless 'classification' is a tautology; no held-out gesture data is used.

  2. other [Section V (Performance Evaluation and Comparison with Prior Work)]
    "we perform noise analysis by adding 1%, 3%, and 5% Gaussian noise to the column current in such a way that the maximum column current decreases by 1%, 3%, and 5% due to noise, while the current of other columns increases by 1%, 3%, and 5% due to noise. ... Overall, we observed97.22% classes can be detected accurately with 5% noise."

    The 'Gaussian noise' is not sampled from a sensor-noise distribution; it is defined as a hand-chosen worst-case perturbation that always lowers the winning column and raises the losing columns. The 97.22% figure is the fraction of these author-constructed scenarios that survive, so the robustness claim is an artifact of the test definition rather than an estimate from noisy data. The result is not a prediction from an independent model but a restatement of the chosen noise scenario.

full rationale

The central classification claim is not derived from first principles: the templates stored as weights are the same waveforms later applied as test inputs, so the near-100% noiseless accuracy is guaranteed by construction. The robustness experiment further defines noise adversarially (max column down, others up), making the 97.22% figure a count over author-chosen scenarios rather than an empirical accuracy. The hardware and energy contributions (0.952 nJ/classification, circuit-level simulations) are independent and not circular, and the self-citations to prior memristor models are used for device parameters rather than to prove the classification result. However, because the paper's headline accuracy is an internal consistency check on a table lookup, not a benchmarked prediction, the circularity score is elevated.

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

The system's claimed accuracy rests on a self-referential test protocol, a hand-chosen voltage-to-resistance mapping, and an idealized simulation of the memristive crossbar and WTA. No independent dataset, device variation, or physical chip results are provided.

free parameters (4)
  • READ threshold (0.45 V) = 0.45 V
    Amplitude threshold deciding whether a sampled sensor value is treated as 3.3 V or 0 V at the READ transistor gate (Sec. IV). Chosen by hand; no justification or study of its effect on accuracy.
  • Sampling schedule (30 ms interval, 8 samples, 2 us window) = 30 ms x 8; 2 us
    The resampled gesture representation used for weight mapping and testing (Sec. I-D and III). The number of samples and interval are chosen without analysis of classification sensitivity.
  • Voltage-to-resistance weight mapping = 5.10 to 96.29 kOhm
    Table II maps each sampled amplitude to a target resistance (SET level). The mapping is not derived from data; intermediate values such as 27.73, 44.69, 78.35 kOhm appear hand-assigned, and this mapping determines the column current margins.
  • Noise scenario levels (1%, 3%, 5%) and direction = 1%/3%/5%
    The robustness test uses contrived noise that always decreases the winning column current and increases the others (Sec. V). The noise model and levels are author-defined, not measured.
assumptions (4)
  • domain assumption HfO2 memristive device behavior follows the Verilog-A model from [9] without statistical variation.
    All crossbar currents are simulated with a single device model in Cadence; no device mismatch, endurance, or temperature variation is included (Sec. II-A, IV).
  • domain assumption The WTA circuit selects the column with maximum current ideally and consumes the stated energy.
    WTA circuit is described verbally (Sec. II-B, IV) with energy numbers but no schematic or simulation; ideal maximum selection is assumed.
  • domain assumption The PV-cell array output combines linearly according to Kirchhoff's voltage law, and motion creates the ternary [0,1,0,-1] patterns.
    The sensing principle (Sec. I-B) assumes algebraic summation of PV voltages and a specific asymmetric matrix kernel; no calibration data is shown beyond a single example per class.
  • ad hoc to paper Incoming sensor signals used for testing are represented by the same eight-sample templates used for programming.
    Table III tests 'if incoming signal = BT' etc. using the same patterns as the stored weights, so the correct column is matched by construction (Sec. IV).

how reviews work

0 comments
Cite this review

Pith. "Pith review of In-Sensor Motion Recognition with Memristive System and Light Sensing Surfaces." pith.science (2026). https://pith.science/paper/LM2XCA7D

@misc{pith2026250606829,
  author       = {Pith},
  title        = {Pith review of: In-Sensor Motion Recognition with Memristive System and Light Sensing Surfaces},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LM2XCA7D}},
  note         = {Machine review of arXiv:2506.06829}
}
read the original abstract

In this paper, we introduce a novel device architecture that merges memristive devices with light-sensing surfaces, for energy-efficient motion recognition at the edge. Our light-sensing surface captures motion data through in-sensor computation. This data is then processed using a memristive system equipped with a HfO2-based synaptic device, coupled with a winner-take-all (WTA) circuit, tailored for low-power motion classification tasks. We validate our end-to-end system using four distinct human hand gestures - left-to-right, right-to-left, bottom-to-top, and top-to-bottom movements - to assess energy efficiency and classification robustness. Our experiments show that the system requires an average of only 4.17 nJ for taking our processed analog signal and mapping weights onto our memristive system and 0.952 nJ for testing per movement class, achieving 97.22% accuracy even under 5% noise interference. A key advantage of our proposed architecture is its low energy requirement, enabling the integration of energy-harvesting solutions such as solar power for sustainable autonomous operation. Additionally, our approach enhances data privacy by processing data locally, reducing the need for external data transmission and storage.

Figures

Figures reproduced from arXiv: 2506.06829 by the authors.

Figure 1
Figure 1. Shows a computational light sensing surface (a) with spatially [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. CMOS-based memristive synapse. Each synapse contains three [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Synaptic crossbar with peripherals to train the system and test the [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages

  1. [1]

    Harper, Ed.,Inside the Smart Home

    R. Harper, Ed.,Inside the Smart Home. London: Springer-Verlag,

    work page

  2. [2]

    People-centric sensing in assistive healthcare: Privacy challenges and directions,

    T. Giannetsos, T. Dimitriou, and N. R. Prasad, “People-centric sensing in assistive healthcare: Privacy challenges and directions,”Secur. Commun. Netw., vol. 4, no. 11, pp. 1295–1307, Nov. 2011

    work page 2011

  3. [3]

    OptoSense: Towards Ubiquitous Self-Powered Ambient Light Sensing Surfaces,

    D. Zhang, J. W. Park, Y . Zhang, Y . Zhao, Y . Wang, Y . Li, T. Bhagwat, W.-F. Chou, X. Jia, B. Kippelen, C. Fuentes-Hernandez, T. Starner, and G. D. Abowd, “OptoSense: Towards Ubiquitous Self-Powered Ambient Light Sensing Surfaces,”Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., vol. 4, no. 3, pp. 1–27, Sep. 2020

    work page 2020

  4. [4]

    SolarGest: Ubiquitous and Battery-free Gesture Recogni- tion using Solar Cells,

    D. Ma, G. Lan, M. Hassan, W. Hu, M. B. Upama, A. Uddin, and M. Youssef, “SolarGest: Ubiquitous and Battery-free Gesture Recogni- tion using Solar Cells,” inMobiCom ’19: The 25th Annual International Conference on Mobile Computing and Networking. New York, NY , USA: Association for Computing Machinery, Aug. 2019, pp. 1–15

    work page 2019

  5. [5]

    Near-sensor and in-sensor computing,

    F. Zhou and Y . Chai, “Near-sensor and in-sensor computing,”Nat. Electron., vol. 3, pp. 664–671, Nov. 2020

    work page 2020

  6. [6]

    Pulse-domain digital image processing for vision chips employing low-voltage operation in deep-submicrometer technologies,

    K. Kagawa, K. Yasuoka, D. C. Ng, T. Furumiya, T. Tokuda, J. Ohta, and M. Nunoshita, “Pulse-domain digital image processing for vision chips employing low-voltage operation in deep-submicrometer technologies,” IEEE J. Sel. Top. Quantum Electron., vol. 10, no. 4, pp. 816–828, Oct. 2004

    work page 2004

  7. [7]

    A 0.5V Real-Time Computational CMOS Image Sensor with Programmable Kernel for Always-On Feature Extraction,

    T.-H. Hsu, Y .-K. Chen, T.-H. Wen, W.-C. Wei, Y .-R. Chen, F.-C. Chang, H. Kim, Q. Chen, B. Kim, R.-S. Liu, C.-C. Lo, K.-T. Tang, M.-F. Chang, and C.-C. Hsieh, “A 0.5V Real-Time Computational CMOS Image Sensor with Programmable Kernel for Always-On Feature Extraction,” in2019 IEEE Asian Solid-State Circuits Conference (A-SSCC). IEEE, 2019, pp. 04–06

    work page 2019

  8. [8]

    Flexible computational photodetectors for self-powered activity sensing,

    D. Zhang, C. Fuentes-Hernandez, R. Vijayan, Y . Zhang, Y . Li, J. W. Park, Y . Wang, Y . Zhao, N. Arora, A. Mirzazadeh, Y . Do, T. Cheng, S. Swaminathan, T. Starner, T. L. Andrew, and G. D. Abowd, “Flexible computational photodetectors for self-powered activity sensing,”npj Flexible Electron., vol. 6, no. 7, pp. 1–8, Jan. 2022

    work page 2022

Show all 27 references
  1. [9]

    Rfam: Reset-failure-aware-model for hfo2-based memristor to enhance the reliability of neuromorphic design,

    H. Das, M. Rathore, R. Febbo, M. Liehr, N. C. Cady, and G. S. Rose, “Rfam: Reset-failure-aware-model for hfo2-based memristor to enhance the reliability of neuromorphic design,” inProceedings of the Great Lakes Symposium on VLSI 2023, ser. GLSVLSI ’23. New York, NY , USA: Asso...

    2023

  2. [10]

    Optimizations for a current-controlled memristor- based neuromorphic synapse design,

    H. Das, R. D. Febbo, C. P. Rizzo, N. N. Chakraborty, J. S. Plank, and G. S. Rose, “Optimizations for a current-controlled memristor- based neuromorphic synapse design,”IEEE Journal on Emerging and Selected Topics in Circuits and Systems, vol. 13, no. 4, pp. 889–900, 2023

    2023

  3. [11]

    An efficient and accurate memristive memory for array-based spiking neural networks,

    H. Das, R. D. Febbo, S. N. B. Tushar, N. N. Chakraborty, M. Liehr, N. C. Cady, and G. S. Rose, “An efficient and accurate memristive memory for array-based spiking neural networks,”IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 70, no. 12, pp. 4804–4815, 2023

    2023

  4. [12]

    Solargest: Ubiquitous and battery-free gesture recognition using solar cells,

    D. Ma, G. Lan, M. Hassan, W. Hu, M. B. Upama, A. Uddin, and M. Youssef, “Solargest: Ubiquitous and battery-free gesture recognition using solar cells,” inThe 25th annual international conference on mobile computing and networking. New York, NY , USA: ACM, 2019, pp. 1– 15

    2019

  5. [13]

    Self-powered gesture recognition with ambient light,

    Y . Li, T. Li, R. A. Patel, X.-D. Yang, and X. Zhou, “Self-powered gesture recognition with ambient light,” inProceedings of the 31st annual ACM symposium on user interface software and technology. New York, NY , USA: ACM, 2018, pp. 595–608

    2018

  6. [14]

    Optosense: Towards ubiquitous self-powered ambient light sensing surfaces,

    D. Zhang, J. W. Park, Y . Zhang, Y . Zhao, Y . Wang, Y . Li, T. Bhagwat, W.-F. Chou, X. Jia, B. Kippelenet al., “Optosense: Towards ubiquitous self-powered ambient light sensing surfaces,”Proceedings of the ACM on interactive, mobile, wearable and ubiquitous technologies, vol....

    2020

  7. [15]

    Flexible computational photodetectors for self-powered activity sensing,

    D. Zhang, C. Fuentes-Hernandez, R. Vijayan, Y . Zhang, Y . Li, J. W. Park, Y . Wang, Y . Zhao, N. Arora, A. Mirzazadehet al., “Flexible computational photodetectors for self-powered activity sensing,”npj Flexible Electronics, vol. 6, no. 1, p. 7, 2022

    2022

  8. [16]

    Neuromorphic motion detection and ori- entation selectivity by volatile resistive switching memories,

    W. Wang, E. Covi, A. Milozzi, M. Farronato, S. Ricci, C. Sbandati, G. Pedretti, and D. Ielmini, “Neuromorphic motion detection and ori- entation selectivity by volatile resistive switching memories,”Advanced Intelligent Systems, vol. 3, no. 4, p. 2000224, 2021

    2021

  9. [17]

    Neuromorphic implementation of a recurrent neural network for emg classification,

    Y . Ma, E. Donati, B. Chen, P. Ren, N. Zheng, and G. Indiveri, “Neuromorphic implementation of a recurrent neural network for emg classification,” in2020 2nd IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS). IEEE, 2020, pp. 69–73

    2020

  10. [18]

    Spike-timing-dependent plasticity for a hafnium-oxide memristive synapse,

    N. N. Chakraborty, H. Das, and G. S. Rose, “Spike-timing-dependent plasticity for a hafnium-oxide memristive synapse,” in2023 IEEE 66th International Midwest Symposium on Circuits and Systems (MWSCAS), 2023, pp. 463–467

    2023

  11. [19]

    Spike-driven synaptic plasticity for a memristive neuromorphic core,

    ——, “Spike-driven synaptic plasticity for a memristive neuromorphic core,” in2023 IEEE 66th International Midwest Symposium on Circuits and Systems (MWSCAS), 2023, pp. 644–648

    2023

  12. [20]

    A mixed-signal short-term plasticity implementation for a current-controlled memristive synapse,

    ——, “A mixed-signal short-term plasticity implementation for a current-controlled memristive synapse,” inProceedings of the Great Lakes Symposium on VLSI 2023, ser. GLSVLSI ’23. New York, NY , USA: Association for Computing Machinery, 2023, p. 179–182

    2023

  13. [21]

    Stdp based online learning for a current-controlled memristive synapse,

    R. Weiss, H. Das, N. N. Chakraborty, and G. S. Rose, “Stdp based online learning for a current-controlled memristive synapse,” in2022 IEEE 65th International Midwest Symposium on Circuits and Systems (MWSCAS), 2022, pp. 1–4

    2022

  14. [22]

    Hardware software co-design for leveraging stdp in a memristive neuroprocessor,

    N. N. Chakraborty, S. O. Ameli, H. Das, C. Schuman, and G. S. Rose, “Hardware software co-design for leveraging stdp in a memristive neuroprocessor,”Neuromorphic Computing and Engineering, 2024. [Online]. Available: http://iopscience.iop.org/article/ 10.1088/2634-4386/ad462b

    2024 doi

  15. [23]

    Homeostatic plasticity in a leaky integrate and fire neuron using tunable leak,

    N. N. Chakraborty, H. Das, and G. S. Rose, “Homeostatic plasticity in a leaky integrate and fire neuron using tunable leak,” in2023 IEEE 66th International Midwest Symposium on Circuits and Systems (MWSCAS), 2023, pp. 738–742

    2023

  16. [24]

    Reliability analysis of memristive reservoir computing architecture,

    M. Rathore, R. Febbo, A. Foshie, S. N. B. Tushar, H. Das, and G. S. Rose, “Reliability analysis of memristive reservoir computing architecture,” inProceedings of the Great Lakes Symposium on VLSI 2023, ser. GLSVLSI ’23. New York, NY , USA: Association for Computing Machinery, ...

    2023

  17. [25]

    Evaluating neuron models through application-hardware co-design,

    C. D. Schuman, H. Das, J. S. Plank, A. Aziz, and G. S. Rose, “Evaluating neuron models through application-hardware co-design,” in2023 57th Asilomar Conference on Signals, Systems, and Computers, 2023

    2023

  18. [26]

    Enhanced read resolution in reconfigurable memristive synapses for spiking neural networks,

    H. Das, C. Schuman, N. N. Chakraborty, and G. S. Rose, “Enhanced read resolution in reconfigurable memristive synapses for spiking neural networks,”Scientific Reports, vol. 14, no. 1, p. 8897, 2024

    2024

  19. [2003]

    Available: http://link.springer.com/10.1007/b97527

    [Online]. Available: http://link.springer.com/10.1007/b97527

    doi

Pith tools

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