Pith. sign in

REVIEW 3 major objections 5 minor 68 references

Adaptive Time-step Training for Enhancing Spike-Based Neural Radiance Fields

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Spike-based NeRF should not render every scene with the same number of time steps; a learned per-scene budget cuts estimated inference energy by up to 68.90% while preserving rendering quality.

desk verdict Scene-wise adaptive time-step learning for spike-based NeRF is a real, well-ablated idea, but the headline energy savings are measured against ANN baselines and overstate PATA's incremental gain by roughly 1.5-2x. read the letter →

arxiv 2507.23033 v2 pith:ZH2QOSZU submitted 2025-07-30 cs.CV cs.NE

Ranxi Lin , Canming Yao , Jiayi Li , Weihang Liu , Xin Lou , Pingqiang Zhou This is my paper
classification cs.CVcs.NE
keywords spikingneuralnetworksradiancefieldsadaptivetimestepscene-wisetemporalbudgetlow-latencyinferenceenergy-efficientrenderingdistillationhybridinputmode
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

Spike-based neural radiance fields render a scene by unrolling a spiking network over a fixed number of time steps, and prior models use the same budget for every scene. The paper argues that NeRF's scene-specific training makes this fixed budget wasteful: simple scenes reach good quality in few steps, while complex scenes need more. It proposes PATA, a two-stage training framework that parameterizes the inference time step as a learnable, scene-wise variable and optimizes it together with rendering quality. The learned budget is shared by all rays in a scene, preserving NeRF's parallel rendering structure while reducing temporal computation. In experiments, the estimated inference energy drops by up to 57.57% on INGP-NeRF and 68.90% on TensoRF, with rendering quality close to full-step and ANN baselines.

What carries the argument

The central object is the learnable scene-wise temporal budget t*_r, defined as Round(Clamp(t*, 1, T)) and optimized with a straight-through estimator. The paper couples this with a hybrid input mode in the spiking neurons: the first time step receives input directly, later steps receive decayed input, so the per-step contribution weight function puts more weight on early steps while still summing to one. A temporal-budget loss then performs the compression, with an extra term penalizing temporal contributions beyond t*_r and a penalty term scaling exp(t*) by the quality ratio M_T / M_smooth(t*), so the budget shrinks only when the target-step output already matches the full-step output. A linear interpolation between errors at the floor and ceiling of t* smooths the discrete rounding, and temporal distillation from the full-step output aligns the shortened rollout with the full-step renderer.

What would settle it

Measure the actual energy of the trained models on a neuromorphic or low-power accelerator while rendering the same scenes at T=8 and at the learned t*_r. If the measured per-scene energy does not fall by roughly the predicted percentages, for example much less than 57.57% on the INGP-NeRF Synthetic-NeRF scenes when the time step drops from 8 to 2.875, then the linear operation-level energy model is doing the work rather than the time-step reduction; alternatively, if rendering PSNR at the reduced step drops by more than about 0.5 dB in a repeated run, the claim that quality is preserved would be falsified.

Watch

Extended reading notes

Core claim

The central claim is that the inference time step of a spike-based NeRF can be treated as a per-scene trainable budget rather than a fixed hyperparameter, and that a two-stage procedure can find a shorter budget without manual search. In the first stage a full-step spike-based renderer is trained at the maximum time step T=8. In the second stage a clamped continuous variable is rounded to an integer target step t*_r, and the network is optimized with a smoothed rendering loss, a temporal-budget loss, and soft supervision from the full-step output. The result is one integer time step per scene, shared by all ray samples; at inference the network unrolls only to t*_r. The authors demonstrate on INGP-NeRF and TensoRF backbones that this preserves parallel rendering and yields estimated energy reductions of up to 57.57% and 68.90%, respectively, with PSNR/SSIM close to the full-step and ANN baselines.

Load-bearing premise

The energy-efficiency results assume inference energy scales linearly with the time step and average firing rate, with fixed per-operation costs of 0.9 pJ per spike operation and 4.6 pJ per float operation; on real hardware with memory, idle, or fixed overheads, the reported percentage savings would not transfer directly.

Editorial extensions

If this is right

  • On INGP-NeRF, increasing the budget-loss weight beta from 1e-7 to 1e-6 moves the learned average time step from 5.125 to 2.875, with PSNR staying between 32.45 and 32.04 dB on Synthetic-NeRF.
  • On TensoRF, the same mechanism works even when only the color decoder is spiking, with estimated energy reductions reaching 40.66% on Synthetic-NeRF and 68.90% on LLFF.
  • The scene-wise shared time step avoids sample-wise or layer-wise early exit, so the method preserves NeRF's parallel rendering pattern instead of breaking it.
  • Because PATA acts on the temporal density and color outputs before volume rendering, it transfers across hash-grid and tensor-factorized representations without changing their input-output interfaces.
  • The learned budgets vary with scene content: scenes with simpler textures use fewer time steps, while scenes with more complex geometry or texture receive larger temporal budgets.

Reading between the lines

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

  • If the linear operation-level energy model is replaced by a measured hardware profile with fixed memory and idle costs, the percentage savings will shrink; the durable quantity to carry forward is the learned time-step reduction itself, not the energy percentage.
  • The same two-stage budget-learning recipe could apply to any temporally-averaged renderer or ensemble, not only spiking networks, since the losses only require time-indexed outputs from the density and color branches.
  • A learned per-scene t*_r also serves as a free complexity signal: the values assigned to complex scenes versus simple scenes could be used by higher-level schedulers to allocate rendering resources across a scene collection.
  • The second training stage still unrolls the network to the full T=8, so PATA does not reduce training cost; a natural extension would be to make the training unroll adaptive as well, which the paper does not attempt.
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

3 major / 5 minor

Summary. The paper proposes PATA, a two-stage training method for spike-based NeRF that learns a scene-wise inference time step t*_r as a trainable variable, combined with a hybrid input mode, full-step pretraining, smoothing, and temporal-budget and distillation losses. The method is evaluated on INGP-NeRF and TensoRF backbones across Synthetic-NeRF, Mip-NeRF 360, and LLFF, reporting competitive PSNR/SSIM with reduced estimated inference energy, including headline reductions of up to 57.57% on INGP-NeRF and 68.90% on TensoRF.

Significance. If the results hold, PATA would be a useful practical contribution: the scene-wise temporal-budget formulation respects NeRF's parallel rendering pattern, the two-stage objective is a sensible way to avoid premature truncation, and the validation across hash-grid and tensor-factorized backbones is a genuine strength. The paper also contains controlled ablations (Tables 5-8) that support the key design choices. However, the headline energy reductions are computed relative to ANN baselines rather than to the fixed full-step spike-based renderer, so the incremental benefit of time-step adaptation is not isolated. Once that baseline is reported, the central claim can be assessed properly; the method itself appears defensible.

major comments (3)
  1. [Tables 1 and 2; Section 5.2.5; Abstract] The energy reductions 'up to 57.57%' and 'up to 68.90%' are computed relative to the ANN baseline, not to the fixed full-step (T=8) spike-based renderer that PATA extends. Because ANN-to-SNN conversion itself removes MAC operations, these numbers conflate conversion gains with time-step-adaptation gains and therefore do not directly support the paper's central claim about learned temporal budgets. Please add the full-step spike-based model to Tables 1 and 2 and recompute all reductions against it. The issue is not cosmetic: under the paper's own Eqs. (22)-(25), the three TensoRF Synthetic-NeRF rows (T=5.625, 3.875, 2.0 with E=885.2, 815.2, 748.4 mJ) imply an approximate full-step spike energy of about 973 mJ, so the reported 40.66% reduction at T=2 becomes about 23% relative to that baseline.
  2. [Section 5.2.5, Eqs. (22)-(25)] The energy estimate assumes an idealized operation-level model: SOPs = fr x T x FLOPs with fixed costs of 0.9 pJ per spike operation and 4.6 pJ per float operation, and no explicit fixed memory, idle, or off-chip cost. The paper does label the numbers as estimates, but all efficiency claims are expressed in these units. Please provide a sensitivity analysis showing how the energy reductions change under alternative per-operation costs and under a fixed per-inference overhead, and state clearly that the time-step reduction is the measured quantity while energy is a modeled extrapolation.
  3. [Tables 1-3] All PSNR/SSIM results appear to be single-run averages without standard deviations or repeated seeds. Several quality differences are within 0.1-0.2 dB (e.g., PATA versus ANN at beta=1e-7 in Table 1), so the claim that PATA 'maintains competitive rendering quality' is not statistically supported. Please report at least three seeds with mean and standard deviation for a representative subset of scenes, or explicitly label the results as single-run exemplars.
minor comments (5)
  1. [Section 4.2, Eq. (8)] The notation w(t,tau) is used in Eq. (8) before it is defined in Eq. (9); please define it at first use and clarify whether O_T is the accumulated membrane potential after T steps or a sum of per-step outputs.
  2. [Section 4.3.3, Eq. (17)] The first term of Eq. (17) uses w_{i,t>t*_r}, but the weight w is not defined with a per-point index i; please define the subscript i or use a different symbol to avoid ambiguity.
  3. [Tables 1 and 2] The tables contain formatting artifacts with numbers running together (e.g., '955.2725.48' and '1.898×10 4'); please use proper column alignment.
  4. [Section 4.3 and Algorithm 1] The second training stage still unrolls the network to the maximum time step T and computes adjacent-step outputs, so total training time and memory cost should be reported alongside inference savings for a complete efficiency picture.
  5. [Table 3] The 'time step' values may not be directly comparable across the listed spike-based NeRF methods because different pipelines define temporal unrolling differently; please state the definition used in each case.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: the learned time step is an explicitly optimized training variable, and the energy reductions follow from the paper's own disclosed energy model rather than from a hidden equivalence.

full rationale

PATA's central claim is not a first-principles derivation. The target time step t* is explicitly parameterized (Eq. 12) and optimized by a loss that contains the temporal-budget penalty L_penalty = sg(M_T/M_smooth(t*)) e^{t*} (Eqs. 16-18), so the direction of the energy effect is built into the training objective rather than discovered post hoc. The reported energy numbers are then arithmetic consequences of Eqs. 22-25, where total energy is linear in the inference time step T. This is a disclosed design-and-evaluation procedure, not a hidden circular equivalence: the non-circular, empirically load-bearing content is that PSNR/SSIM remain competitive (Tables 1, 2, 4), that the learned budget is stable against a fixed-time-step oracle (Table 8), and that quality is evaluated on standard novel-view benchmarks. Refs [26] and [30] include current coauthors, but they are used only as background and as comparison baselines; no uniqueness theorem, ansatz, or core assumption is imported from them. The tendency to report energy reductions relative to ANN baselines rather than to a fixed-full-step spike-based baseline is a baseline-attribution weakness, not a circularity weakness. Overall, no step reduces the paper's conclusion to its inputs by construction; score 2 reflects only the minor, non-load-bearing overlapping self-citations.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard SNN/NeRF machinery plus a set of hand-chosen hyperparameters (notably beta and T=8). The most consequential unverified input is the energy model, which is an estimate imported from prior literature. No new physical entities are introduced.

free parameters (5)
  • beta (temporal-budget loss weight) = 1e-7, 5e-7, 1e-6 (INGP); 5e-9, 5e-8, 5e-7 (TensoRF)
    This weight controls the strength of the compression penalty (Eq. 18) and directly determines the learned time steps and energy reductions reported in Tables 1 and 2. It is tuned by hand to produce different quality-efficiency operating points.
  • alpha (extra-loss weight) = 1e-6
    Fixed hyperparameter weighting the temporal-tail penalty in Eq. 16-17. Chosen without sensitivity analysis.
  • Max time step T = 8
    The full temporal budget used in both training stages and as the reference for energy reduction. The reported time steps and energy savings are relative to this choice.
  • Cauchy loss weights gamma_0, gamma_1, gamma_2 = 5e-5, 5e-2, 1e-6
    Hand-selected weights in the first-stage loss (Eq. 13-15). Not ablated in the paper.
  • PLIF initial decay tau and firing threshold = tau=2.0, threshold=0.5 (TensoRF)
    Initial membrane time constant and threshold for the spiking neurons; chosen by hand and not reported as sensitive.
assumptions (4)
  • domain assumption Volume rendering equation (Eq. 3-4) from NeRF [33] is adopted without modification.
    The paper builds its rendering pipeline on the standard differentiable volume rendering formulation and inherits its assumptions about transmittance and alpha compositing.
  • domain assumption The LIF/PLIF neuron dynamics and surrogate gradient (Eq. 5-7) are valid for training the spiking renderer.
    Standard SNN machinery cited to [13,14,48]; the surrogate gradient is a known approximation whose fidelity is not examined here.
  • domain assumption The operation-level energy model (Eq. 22-25) with 0.9 pJ per SOP and 4.6 pJ per FLOP reflects real inference cost.
    Adopted from prior SNN studies [39,52]; this linear scaling is the basis of the headline energy reductions and is not validated against hardware measurements.
  • domain assumption Temporal information capacity of SNNs increases with time step (citations [8,22]).
    Used to justify that a larger time step can improve rendering quality, motivating the search for a scene-specific budget.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive Time-step Training for Enhancing Spike-Based Neural Radiance Fields." pith.science (2026). https://pith.science/paper/ZH2QOSZU

@misc{pith2026250723033,
  author       = {Pith},
  title        = {Pith review of: Adaptive Time-step Training for Enhancing Spike-Based Neural Radiance Fields},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZH2QOSZU}},
  note         = {Machine review of arXiv:2507.23033}
}
read the original abstract

Spiking Neural Networks (SNNs) provide an energy-efficient computing paradigm for neural rendering, but existing spike-based Neural Radiance Field (NeRF) models usually use a fixed inference time step for all scenes. This fixed temporal budget is inefficient because NeRF follows a scene-specific training paradigm, and different scenes require different temporal capacities to preserve rendering quality. This paper proposes Pretraining-based Adaptive Time-step Adjustment (PATA), a scene-wise adaptive time-step training framework for spike-based NeRF. PATA parameterizes the target inference time step as a trainable variable and optimizes it through a two-stage training process. A hybrid input mode strengthens early time-step outputs, while full-step soft supervision, smoothed rendering loss, and temporal-budget loss jointly maintain rendering fidelity and reduce temporal computation. The learned target time step is shared by all ray samples within a scene, preserving the parallel rendering structure of NeRF. Experiments on INGP-NeRF and TensoRF backbones across Synthetic-NeRF, Mip-NeRF 360, and LLFF show that PATA consistently reduces inference cost while maintaining competitive rendering quality. PATA reduces the estimated inference energy by up to 57.57\% on INGP-NeRF and 68.90\% on TensoRF, demonstrating its effectiveness across different neural rendering representations.

Figures

Figures reproduced from arXiv: 2507.23033 by the authors.

Figure 1
Figure 1. PSNR increases with the inference time step across representative scenes, while different scenes reach comparable quality at different temporal budgets. NeRF should answer two questions: how to learn a scene￾specific temporal budget without manual search, and whether the learned strategy remains effective when the underlying neural rendering representation changes. To address these questions, we propose Pretraining￾… view at source ↗
Figure 2
Figure 2. Overview of the proposed PATA framework with INGP-NeRF as the illustrative backbone. (a) Spike-based neural rendering replaces the nonlinear activations in the density and color MLPs with spiking neurons, producing time-indexed density 𝜎 and color 𝐜 for volume rendering. (b) PATA first trains a full-step renderer at the maximum time step 𝑇 , then optimizes a learnable target time step 𝑡 ∗ using smoothed rendering, d… view at source ↗
Figure 3
Figure 3. Contribution distribution over time steps under the hybrid input mode. (a) Per-step contribution weight. (b) Cumulative contribution up to each time step. The hybrid mode assigns a larger contribution to the first time step and reduces the dominance of late time steps. function for the density network and sigmoid activation for the color network. For the density branch, the activation is applied to the density outpu… view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: further compares GT images, ANN-TensoRF renderings, and PATA-TensoRF renderings on LLFF. In this setting, PATA replaces only the TensoRF color decoder with the spike-based adaptive decoder, while the tensor￾factorized density field, appearance field, sampling process, …
Figure 6
Figure 6. Figure 6: Relationship between PSNR and target time steps [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 62 canonical work pages

  1. [1]

    Mip-nerf: A multiscale representation for anti-aliasingneuralradiancefields,in:ProceedingsoftheIEEE/CVF international conference on computer vision, pp

    Barron,J.T.,Mildenhall,B.,Tancik,M.,Hedman,P.,Martin-Brualla, R., Srinivasan, P.P., 2021. Mip-nerf: A multiscale representation for anti-aliasingneuralradiancefields,in:ProceedingsoftheIEEE/CVF international conference on computer vision, pp. 5855–5864

    work page 2021

  2. [2]

    Barron,J.T.,Mildenhall,B.,Verbin,D.,Srinivasan,P.P.,Hedman,P.,

    work page

  3. [3]

    Zip-nerf: Anti-aliased grid-based neural radiance fields, in: ProceedingsoftheIEEE/CVFInternationalConferenceonComputer Vision, pp

    Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P., 2023. Zip-nerf: Anti-aliased grid-based neural radiance fields, in: ProceedingsoftheIEEE/CVFInternationalConferenceonComputer Vision, pp. 19697–19705

    work page 2023

  4. [4]

    Sah- nerf: Enhancing nerf on novel view synthesis with an snn-ann hybrid framework

    Chang, Y., Peng, P., Zhou, Z., Wang, X., Tian, Y., 2026. Sah- nerf: Enhancing nerf on novel view synthesis with an snn-ann hybrid framework. IEEE Transactions on Circuits and Systems for Video Technology

    work page 2026

  5. [5]

    Tensorf: Ten- sorial radiance fields, in: European conference on computer vision, Springer

    Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H., 2022. Tensorf: Ten- sorial radiance fields, in: European conference on computer vision, Springer. pp. 333–350

    work page 2022

  6. [6]

    How far can we compress instant-ngp-based nerf?, in: Proceedings of the IEEE/CVF ConferenceonComputerVisionandPatternRecognition,pp.20321– 20330

    Chen, Y., Wu, Q., Harandi, M., Cai, J., 2024. How far can we compress instant-ngp-based nerf?, in: Proceedings of the IEEE/CVF ConferenceonComputerVisionandPatternRecognition,pp.20321– 20330

    work page 2024

  7. [7]

    Sres-nerf: improved neuralradiancefieldsforrealismandaccuracyofspecularreflections, in: International Conference on Multimedia Modeling, Springer

    Dai, S., Cao, Y., Duan, P., Chen, X., 2023. Sres-nerf: improved neuralradiancefieldsforrealismandaccuracyofspecularreflections, in: International Conference on Multimedia Modeling, Springer. pp. 306–317

    work page 2023

  8. [8]

    Candeepneuralnetworksbeconverted to ultra low-latency spiking neural networks?, in: 2022 Design, Au- tomation & Test in Europe Conference & Exhibition (DATE), IEEE

    Datta,G.,Beerel,P.A.,2022. Candeepneuralnetworksbeconverted to ultra low-latency spiking neural networks?, in: 2022 Design, Au- tomation & Test in Europe Conference & Exhibition (DATE), IEEE. pp. 718–723

    work page 2022

Show all 68 references
  1. [9]

    Datta, G., Liu, Z., Li, A., Beerel, P.A., 2025. Dynamic spikformer: Low-latency & energy-efficient spiking neural networks with dy- namic time steps for vision transformers, in: ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP),...

    2025

  2. [10]

    JaxNeRF: an ef- ficient JAX implementation of NeRF

    Deng, B., Barron, J.T., Srinivasan, P.P., 2020. JaxNeRF: an ef- ficient JAX implementation of NeRF. URL:https://github.com/ google-research/google-research/tree/master/jaxnerf

    2020

  3. [11]

    Auditory perception architecturewithspikingneuralnetworkandimplementationonfpga

    Deng, B., Fan, Y., Wang, J., Yang, S., 2023. Auditory perception architecturewithspikingneuralnetworkandimplementationonfpga. Neural Networks 165, 31–42. doi:10.1016/j.neunet.2023.05.026

    2023 doi

  4. [12]

    Temporal efficient training of spiking neural network via gradient re-weighting

    Deng, S., Li, Y., Zhang, S., Gu, S., 2022. Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946

    2022 arXiv

  5. [13]

    Spikingjelly: An open-source machinelearninginfrastructureplatformforspike-basedintelligence

    Fang,W.,Chen,Y.,Ding,J.,Yu,Z.,Masquelier,T.,Chen,D.,Huang, L., Zhou, H., Li, G., Tian, Y., 2023. Spikingjelly: An open-source machinelearninginfrastructureplatformforspike-basedintelligence. Science Advances 9, eadi1480

    2023

  6. [14]

    Fang, W., Yu, Z., Chen, Y., Masquelier, T., Huang, T., Tian, Y.,

  7. [15]

    Plenoxels: Radiance fields without neural networks,in:ProceedingsoftheIEEE/CVFConferenceonComputer Vision and Pattern Recognition, pp

    Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A., 2022. Plenoxels: Radiance fields without neural networks,in:ProceedingsoftheIEEE/CVFConferenceonComputer Vision and Pattern Recognition, pp. 5501–5510

    2022

  8. [16]

    Shacira: Scalable hash- grid compression for implicit neural representations, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Girish, S., Shrivastava, A., Gupta, K., 2023. Shacira: Scalable hash- grid compression for implicit neural representations, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 17513–17524

    2023

  9. [17]

    Sharpening your density fields: Spiking neuron aided fast geometry learning

    Gu, Y., Wang, Z., Ye, D., Xu, R., 2024. Sharpening your density fields: Spiking neuron aided fast geometry learning. arXiv preprint arXiv:2412.09881

    2024 arXiv

  10. [18]

    Spike-nerf: Neural radiance field based on spike camera, in: 2024 IEEE International Conference on Multimedia and Expo (ICME), IEEE

    Guo, Y., Bai, Y., Hu, L., Liu, M., Guo, Z., Ma, L., Huang, T., 2024. Spike-nerf: Neural radiance field based on spike camera, in: 2024 IEEE International Conference on Multimedia and Expo (ICME), IEEE. pp. 1–6

    2024

  11. [19]

    Hasssan,A.,Anupreetham,A.,Meng,J.,Seo,J.s.,2025. Quant-nerf: Efficient end-to-end quantization of neural radiance fields with low- precision 3d gaussian representation, in: ICASSP 2025-2025 IEEE InternationalConferenceonAcoustics,SpeechandSignalProcessing (ICASSP), IEEE. pp. 1–5

    2025

  12. [20]

    G-nelf: Memory-and data-efficient hy- brid neural light field for novel view synthesis

    Jiang, L., Wang, L., 2024. G-nelf: Memory-and data-efficient hy- brid neural light field for novel view synthesis. arXiv preprint arXiv:2409.05617

    2024 arXiv

  13. [21]

    Spikingneuralnetworksonfpga:A survey of methodologies and recent advancements

    Karamimanesh, M., Abiri, E., Shahsavari, M., Hassanli, K., van Schaik,A.,Eshraghian,J.,2025. Spikingneuralnetworksonfpga:A survey of methodologies and recent advancements. Neural Networks Lin et al.:Preprint submitted to ElsevierPage 12 of 14 Adaptive Time-step Training for Sp...

    2025

  14. [22]

    Kim, Y., Li, Y., Park, H., Venkatesha, Y., Hambitzer, A., Panda, P.,

  15. [23]

    Unleashing the potential of spiking neural networks with dynamic confidence, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Li, C., Jones, E.G., Furber, S., 2023a. Unleashing the potential of spiking neural networks with dynamic confidence, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 13350–13360

  16. [24]

    Seenn:towardstemporal spiking early exit neural networks

    Li,Y.,Geller,T.,Kim,Y.,Panda,P.,2023b. Seenn:towardstemporal spiking early exit neural networks. Advances in Neural Information Processing Systems 36, 63327–63342

  17. [25]

    Input-aware dynamic timestep spiking neural networks for efficient in-memory computing, in: 2023 60th ACM/IEEE Design Automation Conference (DAC), IEEE

    Li, Y., Moitra, A., Geller, T., Panda, P., 2023c. Input-aware dynamic timestep spiking neural networks for efficient in-memory computing, in: 2023 60th ACM/IEEE Design Automation Conference (DAC), IEEE. pp. 1–6

    2023

  18. [26]

    Spiking-nerf: Spiking neural network for energy-efficient neural rendering

    Li, Z., Ma, Y., Zhou, J., Zhou, P., 2025. Spiking-nerf: Spiking neural network for energy-efficient neural rendering. ACM Journal on Emerging Technologies in Computing Systems 20, 1–23

    2025

  19. [27]

    Im-lif: Improved neuronal dynamics with attention mechanism for direct training deep spiking neural network

    Lian, S., Shen, J., Wang, Z., Tang, H., 2024. Im-lif: Improved neuronal dynamics with attention mechanism for direct training deep spiking neural network. IEEE Transactions on Emerging Topics in Computational Intelligence 8, 2075–2085

    2024

  20. [28]

    Spiking nerf: Repre- senting the real-world geometry by a discontinuous representation, in: Proceedings of the AAAI Conference on Artificial Intelligence, pp

    Liao, Z., Liu, Y., Zheng, Q., Pan, G., 2024. Spiking nerf: Repre- senting the real-world geometry by a discontinuous representation, in: Proceedings of the AAAI Conference on Artificial Intelligence, pp. 13790–13798

    2024

  21. [29]

    Teas: Exploiting spiking activity for temporal-wise adaptive spiking neural networks, in: 2024 29th Asia and South Pacific Design Automation Conference (ASP-DAC), IEEE

    Liu, F., Li, H., Yang, N., Wang, Z., Yang, T., Jiang, L., 2024a. Teas: Exploiting spiking activity for temporal-wise adaptive spiking neural networks, in: 2024 29th Asia and South Pacific Design Automation Conference (ASP-DAC), IEEE. pp. 842–847

    2024

  22. [30]

    Content-aware radi- ance fields: Aligning model complexity with scene intricacy through learnedbitwidthquantization,in:EuropeanConferenceonComputer Vision, Springer

    Liu, W., Zheng, X.X., Yu, J., Lou, X., 2024b. Content-aware radi- ance fields: Aligning model complexity with scene intricacy through learnedbitwidthquantization,in:EuropeanConferenceonComputer Vision, Springer. pp. 239–256

  23. [31]

    Spalen: Sp arsity a ware l oad balancing inference e ngine for neural network

    Longchar, I., Kapoor, H.K., 2024. Spalen: Sp arsity a ware l oad balancing inference e ngine for neural network. ACM Journal on Emerging Technologies in Computing Systems

    2024

  24. [32]

    Local light field fusion: Practicalviewsynthesiswithprescriptivesamplingguidelines

    Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A., 2019. Local light field fusion: Practicalviewsynthesiswithprescriptivesamplingguidelines. ACM Transactions on Graphics 38. doi:10.1145/3306346.3322980

    2019

  25. [33]

    Nerf: Representing scenes as neural radiance fields for view synthesis

    Mildenhall,B.,Srinivasan,P.P.,Tancik,M.,Barron,J.T.,Ramamoor- thi, R., Ng, R., 2021. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65, 99–106

    2021

  26. [34]

    Instant neural graphics primitives with a multiresolution hash encoding

    Müller, T., Evans, A., Schied, C., Keller, A., 2022. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG) 41, 1–15

    2022

  27. [35]

    Surrogate gradient learn- inginspikingneuralnetworks:Bringingthepowerofgradient-based optimization to spiking neural networks

    Neftci, E.O., Mostafa, H., Zenke, F., 2019. Surrogate gradient learn- inginspikingneuralnetworks:Bringingthepowerofgradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine 36, 51–63

    2019

  28. [36]

    Self-architecturalknowledgedistillationfor spikingneuralnetworks

    Qiu, H., Ning, M., Song, Z., Fang, W., Chen, Y., Sun, T., Ma, Z., Yuan,L.,Tian,Y.,2024. Self-architecturalknowledgedistillationfor spikingneuralnetworks. NeuralNetworks178,106475. doi:10.1016/ j.neunet.2024.106475

    2024

  29. [37]

    Exploring neuromorphic computing based on spiking neural networks: Algorithms to hardware

    Rathi, N., Chakraborty, I., Kosta, A., Sengupta, A., Ankit, A., Panda, P., Roy, K., 2023. Exploring neuromorphic computing based on spiking neural networks: Algorithms to hardware. ACM Computing Surveys 55, 1–49

    2023

  30. [38]

    Binary opacity grids:Capturingfinegeometricdetailformesh-basedviewsynthesis

    Reiser,C.,Garbin,S.,Srinivasan,P.,Verbin,D.,Szeliski,R.,Milden- hall, B., Barron, J., Hedman, P., Geiger, A., 2024. Binary opacity grids:Capturingfinegeometricdetailformesh-basedviewsynthesis. ACM Transactions on Graphics (TOG) 43, 1–14

    2024

  31. [39]

    Spikingresformer:bridgingresnetand vision transformer in spiking neural networks, in: Proceedings of the IEEE/CVFConferenceonComputerVisionandPatternRecognition, pp

    Shi,X.,Hao,Z.,Yu,Z.,2024. Spikingresformer:bridgingresnetand vision transformer in spiking neural networks, in: Proceedings of the IEEE/CVFConferenceonComputerVisionandPatternRecognition, pp. 5610–5619

    2024

  32. [40]

    Binary radiance fields

    Shin, S., Park, J., 2023. Binary radiance fields. Advances in neural information processing systems 36, 55919–55931

    2023

  33. [41]

    Scene representa- tion networks: Continuous 3d-structure-aware neural scene represen- tations, in: Advances in Neural Information Processing Systems

    Sitzmann, V., Zollhoefer, M., Wetzstein, G., 2019. Scene representa- tion networks: Continuous 3d-structure-aware neural scene represen- tations, in: Advances in Neural Information Processing Systems

    2019

  34. [42]

    Variablebitrateneuralfields,in:ACM SIGGRAPH 2022 Conference Proceedings, pp

    Takikawa, T., Evans, A., Tremblay, J., Müller, T., McGuire, M., Jacobson,A.,Fidler,S.,2022. Variablebitrateneuralfields,in:ACM SIGGRAPH 2022 Conference Proceedings, pp. 1–9

    2022

  35. [43]

    Nerfstudio: A modular framework for neural radiance field development, in: ACM SIGGRAPH 2023 conference proceedings, pp

    Tancik,M.,Weber,E.,Ng,E.,Li,R.,Yi,B.,Wang,T.,Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., et al., 2023. Nerfstudio: A modular framework for neural radiance field development, in: ACM SIGGRAPH 2023 conference proceedings, pp. 1–12

    2023

  36. [44]

    Torch-ngp: a pytorch implementation of instant-ngp

    Tang, J., 2022. Torch-ngp: a pytorch implementation of instant-ngp. Https://github.com/ashawkey/torch-ngp

    2022

  37. [45]

    Deep learning in spiking neural networks

    Tavanaei, A., Ghodrati, M., Kheradpisheh, S.R., Masquelier, T., Maida, A., 2019. Deep learning in spiking neural networks. Neural networks 111, 47–63

    2019

  38. [46]

    Spik-nerf: Spiking neural networks for neural radiance fields

    Wan, G., Lan, Q., Li, Z., Wang, H., Yitian, W., Li, W., Guo, Y., et al., 2026. Spik-nerf: Spiking neural networks for neural radiance fields. Advances in Neural Information Processing Systems 38, 47316–47335

    2026

  39. [47]

    Wang,K.,Zou,L.,2025.Mixedspikingnerf:Towardsamoreefficient neural radiance fields, in: ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE. pp. 1–5

    2025

  40. [48]

    Spatio-temporalback- propagation for training high-performance spiking neural networks

    Wu,Y.,Deng,L.,Li,G.,Zhu,J.,Shi,L.,2018. Spatio-temporalback- propagation for training high-performance spiking neural networks. Frontiers in neuroscience 12, 331

    2018

  41. [49]

    Hollownerf:Pruning hashgrid-based nerfs with trainable collision mitigation, in: Proceed- ingsoftheIEEE/CVFInternationalConferenceonComputerVision, pp

    Xie,X.,Gherardi,R.,Pan,Z.,Huang,S.,2023. Hollownerf:Pruning hashgrid-based nerfs with trainable collision mitigation, in: Proceed- ingsoftheIEEE/CVFInternationalConferenceonComputerVision, pp. 3480–3490

    2023

  42. [50]

    Back- propagation with sparsity regularization for spiking neural network learning

    Yan, Y., Chu, H., Jin, Y., Huan, Y., Zou, Z., Zheng, L., 2022. Back- propagation with sparsity regularization for spiking neural network learning. Frontiers in Neuroscience 16, 760298

    2022

  43. [51]

    Yang, G.W., Zhou, W.Y., Peng, H.Y., Liang, D., Mu, T.J., Hu, S.M.,

  44. [52]

    Spikingnerf: Making bio-inspired neural networks see through the real world

    Yao, X., Hu, Q., Liu, T., Mo, Z., Zhu, Z., Zhuge, Z., Cheng, J., 2023. Spikingnerf: Making bio-inspired neural networks see through the real world. arXiv preprint arXiv:2309.10987

    2023 arXiv

  45. [53]

    Towards efficient and accurate spiking neural networks via adaptive bit allocation

    Yao, X., Hu, Q., Zhou, F., Liu, T., Li, G., Wang, P., Cheng, J., 2026. Towards efficient and accurate spiking neural networks via adaptive bit allocation. Neural Networks 196, 108350. doi:10.1016/j.neunet. 2025.108350

    2026

  46. [54]

    Glif: A unified gated leaky integrate-and-fire neuron for spiking neural networks

    Yao, X., Li, F., Mo, Z., Cheng, J., 2022. Glif: A unified gated leaky integrate-and-fire neuron for spiking neural networks. Advances in Neural Information Processing Systems 35, 32160–32171

    2022

  47. [55]

    IEEE Transactions on Visualization and Computer Graphics 29, 5124–5136

    Recursive-nerf: An efficient and dynamically growing nerf. IEEE Transactions on Visualization and Computer Graphics 29, 5124–5136

  48. [56]

    Efficient distillation of deep spiking neural networks for full-range timestep deployment

    Yu,C.,Zhao,X.,Liu,L.,Yang,S.,Wang,G.,Li,E.,Wang,A.,2025. Efficient distillation of deep spiking neural networks for full-range timestep deployment. arXiv preprint arXiv:2501.15925

    2025 arXiv

  49. [57]

    Nerf++: Analyzing and improving neural radiance fields

    Zhang, K., Riegler, G., Snavely, N., Koltun, V., 2020. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492

    2020 arXiv

  50. [58]

    Tc-lif: A two-compartment spiking neuron model for long-term sequential modelling, in: Proceedings of the AAAI conference on artificial intelligence, pp

    Zhang, S., Yang, Q., Ma, C., Wu, J., Li, H., Tan, K.C., 2024. Tc-lif: A two-compartment spiking neuron model for long-term sequential modelling, in: Proceedings of the AAAI conference on artificial intelligence, pp. 16838–16847

    2024

  51. [59]

    Accurate and efficient time-domain classification with adaptive spiking recurrent neural networks

    Yin, B., Corradi, F., Bohté, S.M., 2021. Accurate and efficient time-domain classification with adaptive spiking recurrent neural networks. Nature Machine Intelligence 3, 905–913

    2021

  52. [60]

    Zhang,X.,Wang,Y.,Fang,S.,Wang,Z.,Zhang,H.,Qi,D.,Ding,W., 2025b. Asfc-nerf: Large-scale scene rendering with adaptive sam- pling and feature-aware compression, in: ICASSP 2025-2025 IEEE InternationalConferenceonAcoustics,SpeechandSignalProcessing (ICASSP), IEEE. pp. 1–5

    2025

  53. [61]

    Zhong, X., Hu, S., Liu, W., Huang, W., Ding, J., Yu, Z., Huang, T., 2024. Towards low-latency event-based visual recognition with hybridstep-wisedistillationspikingneuralnetworks,in:Proceedings ofthe32ndACMInternationalConferenceonMultimedia,pp.9828– 9836

    2024

  54. [62]

    Gradient aware adaptive quantization: Locally uniform quantization with learnable clipping thresholds for globally non-uniform weights

    Zhou, K., Qiu, Y., Li, Y., Zhao, Q., Zhou, G., 2025. Gradient aware adaptive quantization: Locally uniform quantization with learnable clipping thresholds for globally non-uniform weights. Neural Net- works , 107778

    2025

  55. [63]

    Zhang, T., Yu, K., Zhang, J., Wang, H., 2025a. Da-lif: Dual adaptive leaky integrate-and-fire model for deep spiking neural networks, in: ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE. pp. 1–5. Lin et al.:Preprint subm...

    2025

  56. [64]

    Self-distillation learning based on temporal-spatial consistency for spiking neural networks

    Zuo, L., Ding, Y., Jing, M., Yang, K., Yu, Y., 2024. Self-distillation learning based on temporal-spatial consistency for spiking neural networks. arXiv preprint arXiv:2406.07862 . Lin et al.:Preprint submitted to ElsevierPage 14 of 14

    2024 arXiv

  57. [67]

    Is vanilla mlp in neural radiance field enough for few-shot view synthesis?, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Zhu, H., He, T., Li, X., Li, B., Chen, Z., 2024. Is vanilla mlp in neural radiance field enough for few-shot view synthesis?, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 20288–20298

    2024

  58. [2021]

    2661– 2671

    Incorporating learnable membrane time constant to enhance learningofspikingneuralnetworks,in:ProceedingsoftheIEEE/CVF International Conference on Computer Vision (ICCV), pp. 2661– 2671

  59. [2022]

    5470–5479

    Mip-nerf 360: Unbounded anti-aliased neural radiance fields, in:ProceedingsoftheIEEE/CVFconferenceoncomputervisionand pattern recognition, pp. 5470–5479

  60. [2023]

    8308–8316

    Exploring temporal information dynamics in spiking neural networks, in: Proceedings of the AAAI Conference on Artificial Intelligence, pp. 8308–8316

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.