Pith. sign in

REVIEW 3 major objections 6 minor 20 references

High-Throughput Blind Co-Channel Interference Cancellation for Edge Devices Using Depthwise Separable Convolutions, Quantization, and Pruning

T0 review · 3 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Two compact U-Net models achieve higher MSE scores than WaveNet and ConvTasNet on blind co-channel interference cancellation while requiring far fewer MACs, and depthwise separable versions cut computation by roughly 60% with little MSE…

desk verdict A competent engineering study of efficient U-Nets for blind CCI, with honest trade-off reporting, but the central edge-device claim rests on an untested MSE-to-BER link. read the letter →

arxiv 2411.12541 v1 pith:OV2OYVHA submitted 2024-11-19 eess.SP

Mostafa Naseri , Eli De Poorter , Ingrid Moerman , H. Vincent Poor , Adnan Shahid This is my paper · ORCID
classification eess.SP
keywords blindco-channelinterferencecancellationU-Netdepthwiseseparableconvolutionquantization-awaretrainingedgedevicessignalseparationLSTMMSEscore
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 tries to show that blind co-channel interference cancellation, which removes unknown interfering signals from a desired transmission, can be done accurately enough by small neural networks that fit on edge devices. It proposes two U-Net variants: one shallow model with an LSTM bottleneck (M1) and one deeper fully convolutional model (M2). The paper reports that both achieve higher MSE scores than the WaveNet and ConvTasNet baselines while needing significantly fewer multiply-accumulate operations (MACs). It further shows that replacing standard convolutions with depthwise separable convolutions cuts MACs by 58.66% for M1 and 61.10% for M2, with only a 0.72% MSE degradation for M1 and a 0.63% MSE improvement for M2. The paper explicitly defers bit-error-rate analysis to future work, so the practical communication benefit rests on the untested assumption that the MSE score translates to better decoding.

What carries the argument

The central mechanism is the depthwise separable convolution, which replaces each standard convolution with a per-channel depthwise convolution followed by a pointwise 1x1 convolution, decoupling spatial filtering from channel mixing to cut MACs and parameters. This is combined with quantization-aware training (QAT) that simulates 8-bit precision during fine-tuning of convolutional layers, reducing memory footprint while preserving most of the full-precision performance. The U-Net encoder-decoder structure with skip connections and GroupNorm provides the base architecture; M1 adds an LSTM bottleneck to capture temporal dependencies, while M2 is fully convolutional for input-size flexibility and better GPU batch scaling.

What would settle it

Decode the reconstructed QPSK signals from M1, M2, and the baselines on the same test data and compare bit-error rates; if M1 or M2 do not achieve lower BER than WaveNet and ConvTasNet, the central performance advantage is not confirmed for real communication.

Watch

Extended reading notes

Core claim

The central claim is that two U-Net convolutional architectures, one with an LSTM at the bottleneck (M1) and one fully convolutional (M2), outperform the established WaveNet and ConvTasNet baselines on the ICASSP 2024 blind co-channel interference cancellation task, achieving MSE scores of 44.07 and 42.92 respectively versus 25.45 for WaveNet and 27.66 for ConvTasNet, while requiring substantially fewer MACs and parameters. The paper also establishes that depthwise separable convolutions reduce computational cost to about half (51.25M MACs for M1(Dw) and 51.23M for M2(Dw)) with negligible MSE change, and that quantization-aware training to 8-bit shrinks model size to under 2.1 MB at a larger MSE cost. These results are presented as evidence that optimized U-Nets can bring blind CCI cancellation to resource-constrained edge devices without sacrificing signal reconstruction quality.

Load-bearing premise

The practical value of the reported MSE scores rests on the assumption that a higher truncated negative log-MSE score automatically means better bit decoding after interference cancellation, but the paper does not test bit-error rate.

Editorial extensions

If this is right

  • Deploying blind CCI cancellation on edge devices becomes feasible: the depthwise models are under 2.7 MB and require roughly 51M MACs, far below the billion-scale MAC counts of the baselines.
  • The fully convolutional M2 model scales to the largest batch sizes on a GPU, reaching up to 800,000 symbols per second, making it the better choice for high-throughput applications.
  • Depthwise separable convolutions can be applied to U-Net signal separation models with minimal MSE change, although models with LSTM bottlenecks may see slight degradation because the LSTM relies on rich feature maps.
  • Quantized models M1(Q) at 1.90 MB and M2(Q) at 2.02 MB fit in the memory of low-precision processors, supporting deployment on 8-bit capable edge hardware.
  • Unstructured pruning is not an effective compression strategy for these tightly tuned models, as even a 25% pruning ratio drops MSE scores below 20 for all proposed variants.

Reading between the lines

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

  • A natural next test would be end-to-end BER evaluation with a real decoder on the ICASSP 2024 test data; if the higher MSE scores do not translate into lower bit-error rates, the practical communication benefit is unproven.
  • The efficiency measurements are taken on server CPU and GPU hardware; actual edge processors such as mobile CPUs or microcontrollers may not realize the full MAC reduction as latency or power savings, especially for depthwise convolutions that are not always well optimized.
  • Because the model is trained across all four interference types without using the type labels, a strong robustness test would be to hold out an entire interference type during training and measure blind generalization.
  • The significant MSE drop from quantization (e.g., 44.07 to 33.10 for M1) might still be acceptable if the absolute signal quality remains above the threshold needed for decoding; BER testing would settle whether the trade-off is worthwhile.
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 / 6 minor

Summary. The paper proposes two U-Net-based models, M1 (with an LSTM bottleneck) and M2 (fully convolutional), for blind co-channel interference cancellation on the ICASSP 2024 RF signal separation challenge dataset. It compares these models against WaveNet and ConvTasNet using a truncated negative-log MSE score, MAC counts, parameter counts, and CPU/GPU inference throughput, and it applies depthwise separable convolutions, quantization-aware training, and pruning to reduce computational cost. The headline results are that M1 and M2 achieve higher MSE scores than the baselines with far fewer MACs, and that depthwise separable variants reduce MACs by 58.66% (M1) and 61.10% (M2) with 0.72% degradation and 0.63% improvement in MSE score, respectively. The paper also reports batch-size scaling on GPUs and pruning sensitivity.

Significance. If the comparisons are reliable, the paper offers a genuinely practical efficiency/performance trade-off for learned blind CCI, with the depthwise separable variants being the strongest contribution: the MAC reductions are arithmetically consistent, and M2(Dw) even improves the MSE score. The paper is honest about several limitations, explicitly deferring BER analysis and reporting steep pruning degradation. However, the practical communication-level claim is not yet established because the evaluation metric is a waveform-regression score, not a decoding metric, and the reported symbol rates are raw inference throughputs rather than decoded symbols. No code or seeded training details are provided, which limits independent verification. The central engineering result, if confirmed with communication metrics and repeated-seed statistics, would be a useful baseline for edge-device blind CCI.

major comments (3)
  1. [Section II-B / Section V-C] The central claim that M1 and M2 provide superior blind CCI for edge-device communication is conditional on the MSE score being a valid proxy for bit-error rate, but the paper explicitly defers BER analysis to future work in Section II-B ('a BER analysis that includes bit decoding is left open for future research') and again in Section V-C. Since the MSE score is a truncated negative-log waveform error, a higher MSE score does not imply a lower BER for QPSK with RRC pulse shaping after matched filtering: residual errors below the decision threshold can leave BER unchanged, while correlated or timing-dependent residuals can change BER at the same MSE. I therefore regard the headline practical claim as unsupported until the authors evaluate BER/SINR after an actual receive filter and symbol decision on the challenge test set, or otherwise calibrate the MSE score against BER.
  2. [Section IV-A2 / Section V-B] The 'symbol rate' metric is described in Section IV-A2 as 'the number of communication symbols that can be decoded when performing interference cancellation in one second,' and Section V-B repeats 'up to 800,000 symbols per second on GPUs.' Since the models output an estimated waveform and no bit decoding or demapping is performed, these figures are raw waveform-estimation throughputs, not decoded communication symbols. This overstates the end-to-end communication throughput. The authors should relabel the metric as symbols processed by the cancellation model and, if the intended claim is communication throughput, provide an end-to-end measurement that includes matched filtering and demapping.
  3. [Section IV-B / Table I / Section IV-D / Table II] All MSE scores in Table I and Table II are single-run values with no confidence intervals or repeated-seed statistics. This matters because the key depthwise claims are small relative differences (0.72% degradation for M1, 0.63% improvement for M2), which may be within training noise. Moreover, Table II shows that 1% pruning already reduces the MSE score from 44.07 to 31.67 for M1 and from 42.92 to 29.62 for M2, a loss far larger than the text's 'general trend of degradation' suggests. The pruning section should report repeated-seed variability and explicitly reconcile the 1% results with the stated goal of evaluating compression.
minor comments (6)
  1. [Section III-A1] The component list says LeakyReLU, but the encoder/decoder code equations and Figure 2 consistently use ReLU; please clarify which activation is actually used.
  2. [Section IV-D] The text describes 'unstructured pruning' that removes individual weights, but the parenthetical says 'pruning by removing filters with the lowest l1-norm'; please decide whether the pruning is weight-level or filter-level and state it consistently.
  3. [Table I] Please specify the input length and inference configuration for which MACs and parameter counts are reported, since M2 is described as input-size independent and the MAC numbers must correspond to a particular segment length.
  4. [Section III-C] The dataset description reserves 50 superframes for testing but does not state whether these are balanced across the four interference types; please specify the test-set composition.
  5. [Throughout] There are several typos, including 'Neural Networkss' in Section I-A and 'disurbed' near the end of Section I; a proofreading pass is needed.
  6. [Figures 4-6] The bubble sizes are said to be proportional to parameter counts, but there is no legend or numeric label; please add a legend or labels so the visual comparison is readable.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the paper is an empirical benchmarking study against external baselines and an external challenge dataset, with efficiency gains measured from architectural definitions rather than derived from the target claims.

full rationale

The paper's central claims are experimental comparisons in Section IV and Table I: MSE scores, MACs, parameter counts, and measured inference symbol rates of the proposed U-Net variants M1 and M2 against the external baselines WaveNet and ConvTasNet, using the external ICASSP 2024 grand challenge dataset [14]. Training and evaluation use a held-out set of 50 superframes, so the reported performance is not a fitted or self-defined prediction. The MAC reductions from depthwise separable convolution follow from the architectural definitions and are computed from the model configurations; they are not fitted to the reported MSE outcomes. The MSE score itself is defined in Section II-B and used consistently for all models, so it does not privilege the proposed architectures by construction. The self-citations ([9], [15]) provide architectural context and prior blind-CCI motivation, but Section III specifies the full architectures and training procedure, and the comparison with WaveNet and ConvTasNet does not require accepting those cited results. The paper explicitly defers BER analysis in Section II-B and Section V-C; this is a limitation on the practical communication-level conclusion, but it is not circularity because the paper only claims higher MSE scores and lower computational complexity, not a derived BER result. No load-bearing step reduces to its own input, so the circularity score is 0.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The central comparison is an empirical benchmark. No theoretical derivation is offered, so the ledger lists the hand-tuned architecture choices, the signal-model assumptions, and the metric assumptions that the reported MSE scores and complexity numbers depend on.

free parameters (8)
  • M1 encoder filter counts = [64, 128, 256]
    Chosen by manual hyperparameter search; central to the performance/complexity trade-off.
  • M1 encoder strides = [1, 2, 2]
    Hand-selected downsampling schedule; affects feature resolution and MACs.
  • M2 encoder filter counts = [64, 128, 128, 128, 128, 128, 128, 128, 128]
    Manually tuned depth and width; no ablation is shown to justify this exact sequence.
  • M2 encoder strides = [1, 2, 2, 2, 2, 2, 2, 2, 2]
    Hand-selected downsampling schedule; affects receptive field and memory use.
  • LSTM hidden size in M1 = 64
    Set by hyperparameter search; bottleneck size affects temporal modeling capacity.
  • Kernel size and GroupNorm groups = kernel 3, groups 4
    Fixed by hand; standard choices, not varied or justified beyond convention.
  • Quantization precision = 8-bit (qint8)
    Selected as a design choice; quantized models lose about 11 points of MSE score (M1) and 11.7 points (M2), so the edge-deployment claim depends on accepting this trade-off.
  • Training steps, batch size, learning rate = 100,000 steps, batch size 2, LR 0.002
    Chosen without ablation; these values affect convergence and the final MSE scores.
assumptions (5)
  • domain assumption The received signal is an additive mixture y = s + b of a QPSK/RRC signal of interest and an unknown interference signal b from one of four predefined datasets.
    Section II-A; the evaluation is limited to this specific synthetic/challenge signal model, so the claim is not about general CCI.
  • domain assumption MSE score with threshold at -50 dB is the right performance metric for interference cancellation quality.
    Section II-B; the paper relies on [14] for this metric and explicitly leaves BER analysis to future work. If MSE score does not track BER, the central performance claim is unverified.
  • domain assumption The reserved 50 superframes form an unbiased test set independent of training.
    Section III-C2; the split is by superframes, but the paper does not report the distribution of interference types in each split or verify no leakage through augmentation.
  • domain assumption A single model trained on all four interference types is sufficient for blind operation across unseen interference.
    Section III-A; no evaluation on interference types outside the four challenge categories, so 'blind' handling is tested only within these four types.
  • standard math Standard deep learning components (GroupNorm, ReLU/LeakyReLU, Adam, cosine annealing) work as expected for 1D RF signals.
    Invoked throughout Section III-A; accepted in the literature, not justified in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of High-Throughput Blind Co-Channel Interference Cancellation for Edge Devices Using Depthwise Separable Convolutions, Quantization, and Pruning." pith.science (2026). https://pith.science/paper/OV2OYVHA

@misc{pith2026241112541,
  author       = {Pith},
  title        = {Pith review of: High-Throughput Blind Co-Channel Interference Cancellation for Edge Devices Using Depthwise Separable Convolutions, Quantization, and Pruning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OV2OYVHA}},
  note         = {Machine review of arXiv:2411.12541}
}
read the original abstract

Co-channel interference cancellation (CCI) is the process used to reduce interference from other signals using the same frequency channel, thereby enhancing the performance of wireless communication systems. An improvement to this approach is blind CCI, which reduces interference without relying on prior knowledge of the interfering signal characteristics. Recent work suggested using machine learning (ML) models for this purpose, but high-throughput ML solutions are still lacking, especially for edge devices with limited resources. This work explores the adaptation of U-Net Convolutional Neural Network models for high-throughput blind source separation. Our approach is established on architectural modifications, notably through quantization and the incorporation of depthwise separable convolution, to achieve a balance between computational efficiency and performance. Our results demonstrate that the proposed models achieve superior MSE scores when removing unknown interference sources from the signals while maintaining significantly lower computational complexity compared to baseline models. One of our proposed models is deeper and fully convolutional, while the other is shallower with a convolutional structure incorporating an LSTM. Depthwise separable convolution and quantization further reduce the memory footprint and computational demands, albeit with some performance trade-offs. Specifically, applying depthwise separable convolutions to the model with the LSTM results in only a 0.72% degradation in MSE score while reducing MACs by 58.66%. For the fully convolutional model, we observe a 0.63% improvement in MSE score with even 61.10% fewer MACs. Overall, our findings underscore the feasibility of using optimized machine-learning models for interference cancellation in devices with limited resources.

Figures

Figures reproduced from arXiv: 2411.12541 by the authors.

Figure 1
Figure 1. Root-Raised-Cosine Pulse Shaping Function [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The architecture of the encoder and decoder blocks used in the models. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Model 1 (M1) utilizes a combination of encoder and decoder blocks [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Scatter plot of MSE score versus MACs for different models. The [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Performance comparison of different models on a CPU in terms of [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Performance comparison of different models on a GPU in terms of [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: This graph demonstrates the increasing inference throughput when [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 18 canonical work pages

  1. [9]

    A U-Net Architecture for Time-Frequency Interference Signal Separation of RF Waveforms,

    M. Naseri, J. Fontaine, I. Moerman, E. De Poorter, and A. Shahid, “A U-Net Architecture for Time-Frequency Interference Signal Separation of RF Waveforms,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP), 2024

    work page 2024

  2. [1]

    Music Source Separation in the Waveform Domain,

    A. Défossez, N. Usunier, L. Bottou, and F. Bach, “Music Source Separation in the Waveform Domain,” arXiv preprint arXiv:1911.13254, 2019

    arXiv 1911

  3. [2]

    A Systematic Review of EEG Source Localization Tech- niques and Their Applications on Diagnosis of Brain Abnormalities,

    S. Asadzadeh, T. Yousefi Rezaii, S. Beheshti, A. Delpak, and S. Meshgini, “A Systematic Review of EEG Source Localization Tech- niques and Their Applications on Diagnosis of Brain Abnormalities,” Journal of Neuroscience Methods , vol. 339, p. 108740, 2020

    work page 2020

  4. [3]

    Analysis of Financial Time Series Morphology with Amuse Algorithm and its Extensions,

    R. Szupiluk, T. Z ˛ abkowski, and T. Sobo´n, “Analysis of Financial Time Series Morphology with Amuse Algorithm and its Extensions,” Acta Physica Polonica A, vol. 129, no. 5, pp. 1018–1022, 2016

    work page 2016

  5. [4]

    Unmixing Methods Based on Nonnegativity and Weakly Mixed Pixels for Astronomical Hyperspectral Datasets,

    A. Boulais, O. Berné, G. Faury, and Y . Deville, “Unmixing Methods Based on Nonnegativity and Weakly Mixed Pixels for Astronomical Hyperspectral Datasets,” Astronomy & Astrophysics, vol. 647, p. A105, 2021

    work page 2021

  6. [5]

    A simple and practical underlay scheme for short-range secondary communication,

    M. S. Ibrahim, P. A. Karakasis, and N. D. Sidiropoulos, “A simple and practical underlay scheme for short-range secondary communication,” IEEE Transactions on Wireless Communications , vol. 21, no. 11, pp. 9990–10004, 2022

    work page 2022

  7. [6]

    Enhanced Non-Preemptive Support of URLLC Using Spread Spectrum Underlay Signalling,

    K. S. Bondada, H. Yang, X. Cheng, D. J. Jakubisin, N. Tripathi, G. Anderson, Y . Yang, and J. H. Reed, “Enhanced Non-Preemptive Support of URLLC Using Spread Spectrum Underlay Signalling,” in Proc. IEEE Mil. Commun. Conf. (MILCOM). , 2023, pp. 298–303

    work page 2023

  8. [7]

    Performance Analysis and Deep Learning Design of Underlay Cognitive NOMA- Based CDRT Networks with Imperfect SIC and Co-Channel Interfer- ence,

    T.-H. Vu, T.-V . Nguyen, D. B. da Costa, and S. Kim, “Performance Analysis and Deep Learning Design of Underlay Cognitive NOMA- Based CDRT Networks with Imperfect SIC and Co-Channel Interfer- ence,” IEEE Transactions on Communications, vol. 69, no. 12, pp. 8159– 8174, 2021

    work page 2021

Show all 20 references
  1. [8]

    A Comprehensive Survey on GNSS Interferences and the Application of Neural Networks for Anti- Jamming,

    K. J. Silva Lorraine and M. Ramarakula, “A Comprehensive Survey on GNSS Interferences and the Application of Neural Networks for Anti- Jamming,” IETE Journal of Research , vol. 69, no. 7, pp. 4286–4305, 2023

    2023

  2. [10]

    Improving Data-Driven RF Signal Separation with SOI- Matched Autoencoders,

    L. Henneke, “Improving Data-Driven RF Signal Separation with SOI- Matched Autoencoders,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP) , 2024

    2024

  3. [11]

    DEMUCS for Data-Driven RF Signal Denoising,

    C. Yapar, F. Jaensch, J. C. Hauffen, F. Pezone, P. Jung, S. K. Dehkordi, and G. Caire, “DEMUCS for Data-Driven RF Signal Denoising,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP) , 2024

    2024

  4. [12]

    Linear Versus Non-Linear Interference Cancellation,

    R. M. Buehrer, S. P. Nicoloso, and S. Gollamudi, “Linear Versus Non-Linear Interference Cancellation,” Journal of Communications and Networks, vol. 1, no. 2, pp. 118–133, 1999

    1999

  5. [13]

    Chapter 12 - Multi-User MIMO,

    B. Clerckx and C. Oestges, “Chapter 12 - Multi-User MIMO,” in MIMO Wireless Networks (Second Edition) , second edition ed., B. Clerckx and C. Oestges, Eds. Oxford: Academic Press, 2013, pp. 419–523

    2013

  6. [14]

    The Data-Driven Radio Frequency Signal Separation Challenge,

    T. Jayashankar, B. Kurien, A. Lancho, G. C. Lee, Y . Polyanskiy, A. Weiss, and G. W. Wornell, “The Data-Driven Radio Frequency Signal Separation Challenge,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP), 2024

    2024

  7. [15]

    Blind Co-Channel Interference Cancellation Using Fast Fourier Convolutions,

    M. Naseri, E. de Poorter, I. Moerman, H. V . Poor, and A. Shahid, “Blind Co-Channel Interference Cancellation Using Fast Fourier Convolutions,” in Proc. 2024 IEEE Vehic. Tech. Conf. - Spring (Recent Results) , Singapore, June 24–27, 2024, to appear

    2024

  8. [16]

    ENS-Unet: End-to-End Noise Suppression U-Net for Brain Tumor Segmentation,

    Z. Meng, Z. Fan, Z. Zhao, and F. Su, “ENS-Unet: End-to-End Noise Suppression U-Net for Brain Tumor Segmentation,” in Proc. 40th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. (EMBC) , 2018, pp. 5886–5889

    2018

  9. [17]

    DNoiseNet: Deep Learning- Based Feedback Active Noise Control in Various Noisy Environments,

    Y .-J. Cha, A. Mostafavi, and S. S. Benipal, “DNoiseNet: Deep Learning- Based Feedback Active Noise Control in Various Noisy Environments,” Engineering Applications of Artificial Intelligence , vol. 121, p. 105971, 2023

    2023

  10. [18]

    Mobilenets: Efficient Convolutional Neural Networks for Mobile Vision Applications,

    A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, and H. Adam, “Mobilenets: Efficient Convolutional Neural Networks for Mobile Vision Applications,” arXiv preprint arXiv:1704.04861, 2017

    2017 arXiv

  11. [19]

    Conv-Tasnet: Surpassing Ideal Time– Frequency Magnitude Masking for Speech Separation,

    Y . Luo and N. Mesgarani, “Conv-Tasnet: Surpassing Ideal Time– Frequency Magnitude Masking for Speech Separation,” IEEE/ACM Trans. Audio, Speech, Lang. Process. , vol. 27, no. 8, pp. 1256–1266, 2019

    2019

  12. [20]

    A method to estimate the energy consumption of deep neural networks,

    T.-J. Yang, Y .-H. Chen, J. Emer, and V . Sze, “A method to estimate the energy consumption of deep neural networks,” in 2017 51st Asilomar Conference on Signals, Systems, and Computers , 2017, pp. 1916–1920

    2017

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.