Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

A scalable event-driven spatiotemporal feature extraction circuit

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A redesigned time-difference encoder circuit cuts transistor-mismatch variability of transmitted charge by 61 percent on average, making dense on-chip event-processing arrays feasible.

desk verdict Solid incremental circuit improvement with a well-designed Monte Carlo comparison, but the high-density scalability claim outruns the silicon evidence. read the letter →

arxiv 2501.10155 v2 pith:ZDNMQW6I submitted 2025-01-17 eess.SP cs.ARcs.ET

Hugh Greatorex , Michele Mastella , Ole Richter , Madison Cotteret , Willian Soares Girão , Ella Janotte , Elisabetta Chicca This is my paper · ORCID
classification eess.SPcs.ARcs.ET
keywords timedifferenceencoderevent-drivenprocessingspatiotemporalfeatureextractionCMOSanalogcircuitdevicemismatchMonteCarloanalysisopticalflowdetectionsubthreshold
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 presents a redesigned CMOS circuit for a time difference encoder (TDE), an event-driven processing element that converts the time between spikes on two input channels into an analog output current and thus a rate of output spikes. The new design places a Differential Pair Integrator in both input blocks, which makes the integration of input events linear. Against the earlier TDE circuit, Monte Carlo simulations over 2,000 circuit instances show an average 61 percent reduction in the coefficient of variation (standard deviation divided by mean) of transmitted charge across simulated time differences. The authors argue that this robustness to transistor mismatch is what makes it possible to place many TDEs on one die for real-time parallel processing of event-camera data, and they demonstrate direction-selective optical flow detection using 100 TDE placements run through a fabricated single-TDE chip.

What carries the argument

The load-bearing element is the TDE synapse itself, a two-block analog circuit in which a facilitatory input starts an exponentially decaying subthreshold voltage and a trigger input converts the current value of that trace into a postsynaptic current, so that instantaneous output firing rate encodes the time difference between the two input events. The redesign's central object is the Differential Pair Integrator (DPI), a standard subthreshold CMOS building block that integrates input currents linearly; the new circuit places one DPI in each block, replacing the earlier circuit's single discharge branch. The paper attributes the improved robustness and linear integration to these modifications to the circuit topology.

What would settle it

Build an array of, say, fifty copies of the new TDE on one die in the same 180 nm fabrication process, deliver identical paired input events with fixed time differences to every copy, and measure the coefficient of variation of the integrated output charge across the array. If that measured spread is not clearly smaller than the spread of the earlier circuit, the claimed mismatch robustness is not present in real silicon; a cheaper check is to rerun the Monte Carlo analysis with a foundry-calibrated process-variation model to see whether the 61 percent average reduction survives.

Watch

Extended reading notes

Core claim

The paper's central claim is that the TDE synapse remains accurate in the presence of device mismatch when both its facilitatory and trigger blocks use the same Differential Pair Integrator structure, rather than the single discharge branch used in the earlier design. With this symmetric arrangement, the circuit linearly integrates input events and keeps the exponential mapping from input time difference to output charge intact. The supporting measurement is an average 61 percent reduction in the coefficient of variation of transmitted charge across simulated time differences, obtained from 2000 Monte Carlo instances of each circuit. The authors take this as evidence that the circuit can be integrated into high-density arrays on CMOS technology and demonstrate the concept on an optical-flow task with event-camera data, using 100 TDE placements run through the fabricated single-TDE chip.

Load-bearing premise

The whole scalability conclusion rests on the assumption that the Monte Carlo transistor-variation model used in the 2000-instance simulations accurately represents how the real 180 nm fabrication process varies across an actual chip, since the only silicon evidence is a single fabricated copy of the circuit.

Editorial extensions

If this is right

  • A dense on-chip array of TDEs becomes practical, because the per-unit charge variability under transistor mismatch is reduced enough that different TDEs encode the same time difference in comparable output activity.
  • Event-driven sensors can be coupled to on-chip TDE arrays to perform real-time spatiotemporal feature extraction, such as optical flow detection, at the low power and low latency of subthreshold analog circuits.
  • The circuit preserves the exponential mapping from input time difference to output charge, so output spike rate remains a readable analog code for the timing of input events.
  • Using a DPI in both blocks makes the integration of input events linear, which improves the predictability of the circuit's output across repeated input patterns.
  • The design uses only conventional CMOS transistors, so it can be manufactured in standard 180 nm processes without specialized devices.

Reading between the lines

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

  • Beyond the paper's silicon trace, a natural stress test would be to fabricate a multi-unit array and compare the measured across-unit spread of transmitted charge with the Monte Carlo prediction.
  • If the simulated mismatch reduction transfers to real silicon, the main variability source in large TDE arrays may shift from transistor mismatch to event-source jitter and wiring, which would change system-level design targets.
  • The same two-integrator balancing idea could generalize to other coincidence-detection or correlation circuits that encode inter-event timing in an analog current.
  • A direct practical extension is to use the circuit with real event-camera data rather than simulated data, measuring optical-flow accuracy as a function of TDE array density.
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 / 4 minor

Summary. The paper presents a modified CMOS Time Difference Encoder (TDE) circuit. The new design replaces the single discharge branch in the facilitatory block of the earlier Milde et al. TDE with a differential-pair integrator, aiming for linear integration of input events and improved robustness to device mismatch. The authors report Monte Carlo simulations in Cadence Spectre showing a 61% average reduction in the coefficient of variation of the charge transmitted by the TDE synapse across time differences, and they demonstrate the circuit on an optical flow task using a single on-chip TDE fed sequentially with simulated event-camera data. The conclusion states that the circuit is 'silicon-verified' and suitable for integration into high-density arrays.

Significance. If the central claim holds, the improved TDE is a useful building block for scalable event-driven neuromorphic processing, since mismatch robustness is a known obstacle for analog subthreshold circuits in dense arrays. The paper's main strength is the comparative Monte Carlo study: 2000 simulated instances of both the old and new circuits over a range of time differences provide a concrete, externally benchmarked estimate of improvement. The silicon implementation and the optical-flow demonstration add practical value. However, the significance of the mismatch-robustness claim depends entirely on the fidelity of the unstated Monte Carlo mismatch model to the real XFAB 180 nm process, and on whether the simulation, rather than silicon data, is allowed to carry the scalability conclusion. The paper does not provide the model parameters, any calibration against measured variation, or multi-TDE array data, so the significance is real but currently qualified.

major comments (4)
  1. [Section III-B, Figures 5 and 6] The Monte Carlo analysis is the load-bearing evidence for the 61% reduction in coefficient of variation, but the paper does not report the statistical mismatch model parameters, the PDK version, the assumed per-transistor threshold-voltage and current-factor standard deviations, or any layout-dependent variation. Without these details, a reader cannot reproduce the simulation or assess whether the 2000 instances represent realistic XFAB 180 nm variation. Please report the mismatch model configuration and, ideally, validate the simulation against measured variability from the fabricated chips.
  2. [Section IV Conclusion vs. Sections II-C and III-A] The claim that the circuit is 'silicon-verified' and 'suitable for integration into high-density array structures' overstates the evidence. The only silicon result is a single TDE trace in Figure 3, which demonstrates functionality but not mismatch robustness. The mismatch improvement is shown only in simulation, and Section II-C explicitly states that with multiple TDEs 'further analysis would be needed to examine the impact of mismatch on task performance.' Please either provide multi-TDE array measurements or temper the conclusion to describe the Monte Carlo result as a simulation-based prediction rather than a demonstrated array-level property.
  3. [Section III-C, Figure 7] The optical flow experiment lacks a quantitative performance metric. The percentage of spikes per orientation in Figure 7 shows a qualitative preference for the direction of motion, but there is no measure of detection accuracy, no comparison with the original Milde et al. circuit on the same task, and no error bars or repeated trials. Since only one TDE is used sequentially, this experiment also cannot demonstrate high-density parallel processing. Please add a quantitative metric (e.g., directional selectivity index or correlation with ground-truth motion) and clarify what claim about scalability the experiment can support.
  4. [Section III-A, Figure 3] The silicon trace in Figure 3 is presented without error bars or repeated measurements. If this trace is intended to support the claim of robustness, single-shot data are insufficient. At minimum, the caption should state the number of repeated measurements and the observed variability, or the figure should be clearly labeled as an illustrative functional trace rather than a variability measurement.
minor comments (4)
  1. [Section III-C] The word 'orientanted' should be 'oriented'.
  2. [Acknowledgments] The author contribution line uses 'CRedit' instead of the standard 'CRediT'.
  3. [Figure 5 caption] The caption states that charge is 'normalized relative to the average charge for a given Δt,' but the axis label is ambiguous about the normalization. Please clarify whether the plotted quantity is charge divided by the mean charge, or another normalization.
  4. [Section II-A] The phrase 'linear integration of input events' is central to the design motivation but is not formally defined. Please state in the text what linearity means here (e.g., output charge proportional to the exponential of the time difference) and how the circuit achieves it relative to the earlier design.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central robustness claim is an empirical Monte Carlo comparison against an external baseline, not a reduction to the paper's own assumptions.

full rationale

The paper's central quantitative claim is the average 61% reduction in coefficient of variation of transmitted charge, obtained by Monte Carlo simulation of both the proposed circuit and the Milde et al. 2018 baseline under the same simulated process variation. This is an external comparison: the result is measured from simulation outputs, not derived from the paper's definitions or from fitted parameters. The bias voltages used are operational settings, not tuned to reproduce the 61% figure. The optical flow demonstration is likewise an empirical test against synthetic event-camera data, with the circuit used as a measurement subject. No equation in the paper defines the improvement into existence, and no load-bearing argument reduces to a self-citation; the citations to the authors' own prior work are background context (e.g., the cognigr1 IC and previous TDE implementations), not the basis for the robustness conclusion. The skeptical concern about whether Cadence Spectre Monte Carlo models faithfully represent real XFAB 180 nm mismatch is a validity or correctness risk, not circularity: it questions whether the simulation evidence transfers to silicon, but it does not make the simulation's output equivalent to the paper's inputs. The paper itself flags the need for further array-level mismatch analysis in Sec. II-C, which is honest rather than circular. Therefore the appropriate circularity score is 0.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No free parameters are fitted to data; circuit biases are configured via on-chip DAC but do not affect the robustness comparison. The central claims rest on the DPI model, the Monte Carlo process variation model, and the synthetic data realism.

assumptions (3)
  • domain assumption Transistor subthreshold operation and DPI circuit behavior as described in Bartolozzi and Indiveri [19] apply to the fabricated 180nm devices.
    The improved circuit's linear integration claim relies on the DPI integrator model, referenced from [19], without fresh silicon validation beyond a single trace.
  • ad hoc to paper The Monte Carlo mismatch distribution used in Cadence Spectre represents realistic physical process variation for XFAB 180nm.
    The paper states 'realistic distribution' but does not specify the statistical parameters; the conclusion of scalability depends on this assumption.
  • domain assumption The synthetic event-based camera data and added jitter adequately model a real AER sensor for the optical flow task.
    Used in Section II-C to validate the application; no real sensor data was used.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A scalable event-driven spatiotemporal feature extraction circuit." pith.science (2026). https://pith.science/paper/ZDNMQW6I

@misc{pith2026250110155,
  author       = {Pith},
  title        = {Pith review of: A scalable event-driven spatiotemporal feature extraction circuit},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZDNMQW6I}},
  note         = {Machine review of arXiv:2501.10155}
}
read the original abstract

Event-driven sensors, which produce data only when there is a change in the input signal, are increasingly used in applications that require low-latency and low-power real-time sensing, such as robotics and edge devices. To fully achieve the latency and power advantages on offer however, similarly event-driven data processing methods are required. A promising solution is the TDE: an event-based processing element which encodes the time difference between events on different channels into an output event stream. In this work we introduce a novel TDE implementation on CMOS. The circuit is robust to device mismatch and allows the linear integration of input events. This is crucial for enabling a high-density implementation of many TDEs on the same die, and for realising real-time parallel processing of the high-event-rate data produced by event-driven sensors.

Figures

Figures reproduced from arXiv: 2501.10155 by the authors.

Figure 1
Figure 1. Photograph of the realised “cognigr1” Application Specific [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The circuit schematics of the TDE synapse. The circuit initially proposed in Milde [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Silicon measurements of the TDE circuit. The TDE received [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Left: Simulated event-based camera data of a textured surface moving in a vertical direction. Center: The x-y space was randomly [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]
Figure 5
Figure 5. Figure 5: Monte Carlo simulation results comparing our circuit [PITH_FULL_IMAGE:figures/full_fig_p003_5.png]
Figure 6
Figure 6. Figure 6: The average charge transmitted plotted against [PITH_FULL_IMAGE:figures/full_fig_p003_6.png]
Figure 7
Figure 7. Figure 7: Left: A raster plot of the measured spiking responses of the [PITH_FULL_IMAGE:figures/full_fig_p004_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Event-based vision for egomotion estimation using precise event timing

    cs.CV 2025-01 reject novelty 5.0 of 10

    A hand-configured spiking network of time-difference encoders estimates vehicle yaw rate from event-camera data, claiming order-of-magnitude lower rotational error than prior event-based methods.

Reference graph

Works this paper leans on

29 extracted references · 12 canonical work pages · cited by 1 Pith paper

  1. [1]

    Activity-driven, event-based vision sensors,

    T. Delbr ¨uck, B. Linares-Barranco, E. Culurciello, and C. Posch, “Activity-driven, event-based vision sensors,” in IEEE ISCAS 2010 , 2010, pp. 2426–2429. DOI: 10.1109/ISCAS.2010.5537149

    work page arXiv 2010

  2. [2]

    Neuromorphic sensing, per- ception and control for robotics,

    C. Bartolozzi, A. Glover, and E. Donati, “Neuromorphic sensing, per- ception and control for robotics,” in Handbook of Neuroengineering . 2020, pp. 1–31. DOI: 10.1007/978-981-15-2848-4 116-1

    work page doi:10.1007/978-981-15-2848-4 2020

  3. [3]

    An analog electronic cochlea,

    R. Lyon and C. Mead, “An analog electronic cochlea,” IEEE Transac- tions on Acoustics, Speech, and Signal Processing , vol. 36, pp. 1119– 1134, 1988. DOI: 10.1109/29.1639

    doi:10.1109/29.1639 1988

  4. [4]

    Spike-based dynamic computing with asynchronous sensing-computing neuromorphic chip,

    M. Yao et al. , “Spike-based dynamic computing with asynchronous sensing-computing neuromorphic chip,” Nature Communications , vol. 15, p. 4464, 2024. DOI: 10.1038/s41467-024-47811-6

    doi:10.1038/s41467-024-47811-6 2024

  5. [5]

    Event-driven sensing for efficient perception: Vision and audition algorithms,

    S.-C. Liu, B. Rueckauer, E. Ceolini, A. Huber, and T. Delbruck, “Event-driven sensing for efficient perception: Vision and audition algorithms,” IEEE Signal Processing Magazine , vol. 36, pp. 29–37,

    work page

  6. [6]

    Event- Based Object Detection and Tracking for Space Situational Aware- ness,

    S. Afshar, A. P. Nicholson, A. van Schaik, and G. Cohen, “Event- Based Object Detection and Tracking for Space Situational Aware- ness,” IEEE Sensors Journal , vol. 20, pp. 15 117–15 132, 2020. DOI: 10.1109/JSEN.2020.3009687

    arXiv 2020

  7. [7]

    Finding the gap: Neuromorphic motion-vision in dense environments,

    T. Schoepe, E. Janotte, M. B. Milde, O. J. N. Bertrand, M. Egelhaaf, and E. Chicca, “Finding the gap: Neuromorphic motion-vision in dense environments,” Nature Communications, vol. 15, p. 817, 2024. DOI: 10.1038/s41467-024-45063-y

    work page doi:10.1038/s41467-024-45063-y 2024

  8. [8]

    Odour localization in neuromorphic systems,

    T. Schoepe et al., “Odour localization in neuromorphic systems,” in IEEE ISCAS 2024 , 2024, pp. 1–5. DOI: 10.1109/ISCAS58744.2024. 10558186

    arXiv 2024

Show all 29 references
  1. [9]

    Closed-loop sound source localization in neuro- morphic systems,

    T. Schoepe et al. , “Closed-loop sound source localization in neuro- morphic systems,” Neuromorphic Computing and Engineering, vol. 3, p. 024 009, 2023. DOI: 10.1088/2634-4386/acdaba

    2023 doi

  2. [10]

    A hardware-friendly neuromorphic spik- ing neural network for frequency detection and fine texture decoding,

    M. Mastella and E. Chicca, “A hardware-friendly neuromorphic spik- ing neural network for frequency detection and fine texture decoding,” in IEEE ISCAS 2021, 2021, pp. 1–5. DOI: 10.1109/ISCAS51556.2021. 9401377

    2021

  3. [11]

    Event-based eccentric motion detection exploit- ing time difference encoding,

    G. D’Angelo et al., “Event-based eccentric motion detection exploit- ing time difference encoding,” Frontiers in Neuroscience , vol. 14,

  4. [12]

    Low-latency monocular depth estimation using event timing on neuromorphic hardware,

    S. Chiavazza, S. M. Meyer, and Y . Sandamirskaya, “Low-latency monocular depth estimation using event timing on neuromorphic hardware,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops , 2023, pp. 4071– 4080

    2023

  5. [13]

    Loihi: A neuromorphic manycore processor with on-chip learning,

    M. Davies et al. , “Loihi: A neuromorphic manycore processor with on-chip learning,” IEEE Micro , vol. 38, pp. 82–99, 2018. DOI: 10. 1109/MM.2018.112130359

    2018

  6. [14]

    A million spiking-neuron integrated circuit with a scalable communication network and interface,

    P. A. Merolla et al., “A million spiking-neuron integrated circuit with a scalable communication network and interface,” Science, vol. 345, pp. 668–673, 2014. DOI: 10.1126/science.1254642

    2014 doi

  7. [15]

    Mead, Analog VLSI and neural systems

    C. Mead, Analog VLSI and neural systems . 1989. DOI: 10.1007/978- 1-4613-1639-8

    1989 doi

  8. [16]

    A silicon neuron,

    M. Mahowald and R. Douglas, “A silicon neuron,” Nature, vol. 354, pp. 515–518, 1991. DOI: 10.1038/354515a0

    1991 doi

  9. [17]

    Event-based vision: A survey,

    G. Gallego et al., “Event-based vision: A survey,” IEEE transactions on pattern analysis and machine intelligence , vol. 44, pp. 154–180,

  10. [18]

    Spiking Elementary Motion Detector in Neuromorphic Systems,

    M. B. Milde, O. J. N. Bertrand, H. Ramachandran, M. Egelhaaf, and E. Chicca, “Spiking Elementary Motion Detector in Neuromorphic Systems,” Neural Computation, vol. 30, pp. 2384–2417, 2018. DOI: 10.1162/neco a 01112

    2018 doi

  11. [19]

    Synaptic dynamics in analog VLSI,

    C. Bartolozzi and G. Indiveri, “Synaptic dynamics in analog VLSI,” Neural Computation , vol. 19, pp. 2581–2603, 2007. DOI: 10.1162/ neco.2007.19.10.2581

    2007

  12. [20]

    DOI: 10.1109/TPAMI.2020.3008413

    2020

  13. [21]

    A reconfigurable neuromor- phic VLSI multi-chip system applied to visual motion computation,

    G. Indiveri, A. Whatley, and J. Kramer, “A reconfigurable neuromor- phic VLSI multi-chip system applied to visual motion computation,” Proceedings of the Seventh International Conference on Microelec- tronics for Neural, Fuzzy and Bio-Inspired Systems , pp. 37–44, 1999. DOI: 1...

    1999

  14. [22]

    Event-based vision for egomotion estimation using precise event timing,

    H. Greatorex, M. Mastella, M. Cotteret, O. Richter, and E. Chicca, “Event-based vision for egomotion estimation using precise event timing,” arXiv, 2025. arXiv: 2501.11554

    2025 arXiv

  15. [23]

    Pulse-based analog VLSI velocity sen- sors,

    J. Krammer and C. Koch, “Pulse-based analog VLSI velocity sen- sors,” IEEE Transactions on Circuits and Systems II: Analog and Digital Signal Processing , vol. 44, pp. 86–101, 1997. DOI: 10.1109/ 82.554431

    1997

  16. [25]

    Robust spiking attractor networks with a hard winner-take-all neuron circuit,

    M. Cotteret et al. , “Robust spiking attractor networks with a hard winner-take-all neuron circuit,” in 2023 IEEE International Sympo- sium on Circuits and Systems (ISCAS) , 2023, pp. 1–5. DOI: 10.1109/ ISCAS46773.2023.10181513

    2023

  17. [26]

    A current-mode conductance-based silicon neuron for address-event neuromorphic systems,

    P. Livi and G. Indiveri, “A current-mode conductance-based silicon neuron for address-event neuromorphic systems,” in IEEE ISCAS 2009, 2009, pp. 2898–2901. DOI: 10.1109/ISCAS.2009.5118408

    2009

  18. [27]

    Communicating neuronal ensembles between neu- romorphic chips,

    K. A. Boahen, “Communicating neuronal ensembles between neu- romorphic chips,” in Neuromorphic Systems Engineering: Neural Networks in Silicon . 1998, pp. 229–259. DOI: 10.1007/978-0-585- 28001-1 11

    1998 doi

  19. [29]

    A subthreshold second-order integration circuit for versatile synaptic alpha kernel and trace generation,

    O. Richter et al., “A subthreshold second-order integration circuit for versatile synaptic alpha kernel and trace generation,” in Proceedings of the 2023 International Conference on Neuromorphic Systems, 2023. DOI: 10.1145/3589737.3606008

    2023

  20. [2019]

    DOI: 10.1109/MSP.2019.2928127

    2019

  21. [2020]

    DOI: 10.3389/fnins.2020.00451

    2020

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.