Back

SCINet: Time Series Modeling and Forecasting with Sample Convolution and Interaction

Minhao LiuAiling ZengMuxi ChenZhijian XuQiuxia LaiLingna MaQiang Xu

article2022NeurIPS933 citations

Proposes a hierarchical downsample-convolve-interact neural network architecture that captures multi-resolution temporal features to outperform existing convolutional and Transformer-based models on complex time series forecasting tasks.

Listen

Accurate time series forecasting is critical for strategic decision-making across industries such as energy management, traffic planning, healthcare, and financial investment. While recurrent neural networks, temporal convolutional networks, and Transformer-based models are commonly used for sequence modeling, they often fail to exploit the inherent structural properties of time series data. Specifically, standard approaches struggle to balance broad temporal context with localized temporal dynamics, resulting in suboptimal predictive performance and high computational complexity.

The article evaluates a new deep learning framework called the Sample Convolution and Interaction Network (SCINet) designed to improve both short-term and long-term time series forecasting. The main objective is to demonstrate that recursively downsampling time series data, extracting features via distinct convolutional filters, and enabling bidirectional interactive learning between sub-sequences produces a more predictable representation with superior forecasting accuracy.

The researchers designed a hierarchical downsample-convolve-interact architecture structured as a binary tree of basic blocks, which can also be stacked with intermediate supervision for complex dynamics. They evaluated the model using 11 public real-world benchmark datasets covering electricity demand, transformer temperatures, solar power, exchange rates, and freeway traffic systems. The evaluation spanned short-term, long-term, multivariate, univariate, and spatial-temporal forecasting tasks against established recurrent, convolutional, and Transformer baselines.

The evaluation revealed several key findings. First, the proposed model consistently outperformed prior state-of-the-art methods, achieving an average 39.89% reduction in mean squared error across benchmark long-term forecasting tasks and up to a 65% reduction in error on exchange rate data. Second, in short-term forecasting, the model improved accuracy over conventional models by up to 10%, while Transformer models performed poorly due to their lack of focus on recent local temporal patterns. Third, in spatial-temporal traffic forecasting benchmarks, the architecture outperformed dedicated graph neural network models across multiple metrics without relying on explicit spatial relation modeling. Fourth, complexity analysis demonstrated that the model scales with a worst-case computational time complexity of O(T log T), making it significantly more efficient than standard attention-based Transformer models that scale quadratically at O(T^2).

These results indicate that specialized downsampling and interaction architectures can substantially lower operational prediction errors while reducing computational costs compared to complex Transformer architectures. This performance-to-compute advantage directly affects resource-constrained operational environments, such as real-time grid balancing or dynamic traffic rerouting, where high-latency models are impractical. Furthermore, the findings challenge the prevailing assumption that large attention mechanisms are necessary for modeling long-range temporal dependencies.

Organizations seeking to optimize sequence forecasting pipelines should consider piloting downsampling-and-interaction convolution architectures as an alternative or complement to Transformer-based and graph-based models. Operational teams should evaluate these architectures on both short and long horizons, particularly where inference cost and latency are constraining factors. However, because this architecture currently targets deterministic, regularly sampled data, decision-makers should exercise caution when deploying it on datasets characterized by severe missing values or irregular sampling intervals, and they should await probabilistic forecasting extensions before relying on it for uncertainty-sensitive risk management.

The Chinese University of Hong Kongconvolutional neural networksencoder-decodernetwork architectureSCINettime series modelingsample convolution
arXiv: 2106.09305cure-lab/SCINet
Cover for SCINet: Time Series Modeling and Forecasting with Sample Convolution and Interaction

Abstract

One unique property of time series is that the temporal relations are largely preserved after downsampling into two sub-sequences. By taking advantage of this property, we propose a novel neural network architecture that conducts sample convolution and interaction for temporal modeling and forecasting, named SCINet. Specifically, SCINet is a recursive downsample-convolve-interact architecture. In each layer, we use multiple convolutional filters to extract distinct yet valuable temporal features from the downsampled sub-sequences or features. By combining these rich features aggregated from multiple resolutions, SCINet effectively models time series with complex temporal dynamics. Experimental results show that SCINet achieves significant forecasting accuracy improvements over both existing convolutional models and Transformer-based solutions across various real-world time series forecasting datasets. Our codes and data are available at https://github.com/cure-lab/SCINet.

Topics

convolutional neural networksencoder-decodernetwork architectureconvolutional codesarchitecture designsfeedforward neural networksforecasting systemforecasting techniquestemporal correlations

Keywords

SCINettime series modelingsample convolutioninteraction networktemporal convolutional networkstraffic forecastingspatial-temporal time seriessequence modelingshort-term forecastingunivariate forecasting

Table of Contents

  • 1 Introduction
  • 2 Related Work and Motivation
  • 2.1 Related Work
  • 2.2 Rethinking Dilated Causal Convolution for Time Series Modeling and Forecasting
  • 3 SCINet: Sample Convolution and Interaction Network
  • 3.1 SCI-Block
  • 3.2 SCINet
  • 3.3 Stacked SCINet
  • 3.4 Loss Function
  • 3.5 Complexity Analysis
  • 4 Experiments
  • 4.1 Datasets
  • 4.2 Results and Analyses
  • 4.3 Ablation studies
  • 5 Limitations and Future Work
  • 6 Conclusion
  • References
  • A Datasets and Evaluation Metrics
  • A.1 Electricity Transformer Temperature (ETT)
  • A.2 PeMS
  • A.3 Traffic, Solar-Energy, Electricity and Exchange-Rate
  • B Extra Experimental Results
  • B.1 Error Bars Evaluation
  • B.2 Evaluation on the Impact of KK and LL
  • B.3 Empirical Study on Operator Selection
  • C Reproducibility

Knowls

  1. Knowl 1 — SCI-Block couples even–odd downsampling with learned cross-interaction

    model/method

    An SCI-Block takes a time-ordered feature sequence FF and separates its even- and odd-indexed elements into subsequences FevenF_{\mathrm{even}} and FoddF_{\mathrm{odd}}. Each subsequence is transformed using learned one-dimensional convolution modules, and the two streams exchange information through a pair of affine-style updates. For aligned feature tensors, the updates are

    Fodds=Fodd⊙exp⁡(ϕ(Feven)),Fevens=Feven⊙exp⁡(ψ(Fodd)),F^{s}_{\mathrm{odd}} = F_{\mathrm{odd}} \odot \exp(\phi(F_{\mathrm{even}})), \qquad F^{s}_{\mathrm{even}} = F_{\mathrm{even}} \odot \exp(\psi(F_{\mathrm{odd}})), Fodd′=Fodds±ρ(Fevens),Feven′=Fevens±η(Fodds).F'_{\mathrm{odd}} = F^{s}_{\mathrm{odd}} \pm \rho(F^{s}_{\mathrm{even}}), \qquad F'_{\mathrm{even}} = F^{s}_{\mathrm{even}} \pm \eta(F^{s}_{\mathrm{odd}}).

    Here, FsF^s denotes a scaled feature, F′F' an updated feature, ⊙\odot elementwise multiplication, and exp⁡\exp an elementwise exponential. The learned convolution modules ϕ,ψ,ρ,η\phi,\psi,\rho,\eta map between compatible feature shapes; they have distinct parameters in the standard block. The scaling of each stream is computed from the other stream, and the subsequent update adds or subtracts a transformed version of the other scaled stream. This interaction is intended to restore information that could otherwise be lost by splitting and downsampling.

  2. Knowl 2 — SCINet forms a multiresolution tree and decodes a residual-enhanced sequence

    model/method

    For an input time series X∈RT×dX\in\mathbb{R}^{T\times d}, where TT is the look-back length and dd the number of variates, SCINet arranges SCI-Blocks in a binary tree. At level ll there are 2l2^l blocks, for levels l=1,…,Ll=1,\ldots,L; each block splits its input into even- and odd-indexed subsequences and processes them with cross-interaction. Repeated splitting gives blocks access to different temporal resolutions, while information from shallower levels is carried into deeper features. After the final level, SCINet reverses the splits to realign the features as a sequence, adds that representation to the original input through a residual connection, and uses a fully connected decoder to produce the forecast X^∈Rτ×d\hat X\in\mathbb{R}^{\tau\times d} for horizon length τ\tau. For tasks where distribution shift is a concern, the model can subtract the final observed value of each variate from the look-back series and add it back to every forecasted value.

  3. Knowl 3 — Stacked SCINets use forecast feedback and supervision at every stage

    model/method

    A stacked SCINet contains KK SCINets. The first receives the length-TT historical input and predicts a horizon of length τ\tau. Each subsequent SCINet receives a length-TT sequence formed by joining the previous stage's τ\tau predictions to the most recent T−τT-\tau observed values. Each stage is supervised against the same ground-truth forecast, and the final stage's output is the system forecast. For stage kk, let x^i(k)\hat x_i^{(k)} be its prediction and xix_i the target vector at horizon step ii, with xi,x^i(k)∈Rdx_i,\hat x_i^{(k)}\in\mathbb{R}^d. The training objective is the sum of the mean horizon-wise L1 losses:

    Lk=1τ∑i=1τ∥x^i(k)−xi∥1,L=∑k=1KLk.\mathcal{L}_k=\frac{1}{\tau}\sum_{i=1}^{\tau}\left\|\hat x_i^{(k)}-x_i\right\|_1, \qquad \mathcal{L}=\sum_{k=1}^{K}\mathcal{L}_k.

  4. Knowl 4 — SCINet achieves broad receptive fields with relatively few levels

    theoretical result

    The paper gives SCINet a worst-case time complexity of O(Tlog⁡T)O(T\log T) for look-back length TT, compared with O(T2)O(T^2) for a vanilla Transformer, and describes its computational cost as usually on par with a temporal convolutional network. The authors attribute SCINet's broad receptive fields to progressive downsampling: each convolutional layer operates on a coarser sequence than the original. In their experiments, forecasting accuracy was typically best with at most L=5L=5 tree levels, including for a look-back length such as T=168T=168, and at most K=3K=3 stacked SCINets was generally sufficient. These are reported empirical design observations, not guarantees for every dataset.

  5. Knowl 5 — Short- and long-horizon benchmarks show improvements over comparison models

    data/table

    On short-term multivariate forecasting, SCINet used a look-back length of 168 and horizons 3,6,12,243,6,12,24. The table gives the paper's reported percentage improvement over the best comparison model for each setting, measured using RSE. On long-term forecasting, the horizons were 96,192,336,72096,192,336,720; the table gives SCINet's MSE and reported MSE improvement over the best comparison model. The short-term gains are positive in all listed settings, while the long-term results show particularly large gains on Exchange Rate. Across the long-term settings, the authors report a 39.89% average MSE improvement and an average improvement of about 65% on Exchange Rate.

    Could not parse LaTeX table

  6. Knowl 6 — Multivariate ETT results favor SCINet at 14 of 15 forecast horizons

    empirical result

    In multivariate forecasting on ETTh1, ETTh2, and ETTm1, SCINet had lower MSE than every listed comparison method at 14 of the 15 dataset–horizon settings. The exception was ETTh1 at horizon 720, where SCINet's MSE was 0.544 and Autoformer's was 0.499. The values below compare SCINet with the lowest MSE among the other models reported for each setting; the ETTm1 horizons are 24, 48, 96, 288, and 672, while the ETTh datasets use 24, 48, 168, 336, and 720. Input lengths were kept the same as those used by Informer. Qualitative forecasts on randomly selected ETTh1 sequences also illustrated SCINet's ability to track trend and seasonal patterns.

    Could not parse LaTeX table

  7. Knowl 7 — SCINet outperforms the reported baselines on all PeMS spatial-temporal metrics

    data/table

    The PeMS experiments evaluate 12-step spatial-temporal forecasts on PEMS03, PEMS04, PEMS07, and PEMS08. SCINet does not explicitly model spatial relations, but its reported MAE, MAPE, and RMSE are lower than those of every comparison method that reports the corresponding metric on each dataset. The table reproduces SCINet's values and the paper's reported percentage improvement over the best comparison for each metric; no units are specified in the reported results.

    Could not parse LaTeX table

  8. Knowl 8 — Ablations support cross-interaction, distinct filters, residuals, and decoding

    empirical result

    Ablations on ETTh1 and PEMS08 compared a complete SCINet with variants that removed cross-interaction, shared the SCI-Block transformation weights, removed the residual connection, or removed the fully connected decoder. The SCI-Block ablations used one SCINet with L=3L=3 levels; the evaluated forecast settings included ETTh1 at horizon 720 and PEMS08 at horizon 12. The plotted mean-absolute-error comparisons showed the complete model performing best among the tested variants. Removing cross-interaction or sharing transformation weights worsened accuracy, supporting both information exchange between subsequences and distinct learned transformations. The authors also report that interaction was more effective with longer look-back windows. Removing the residual connection caused a substantial drop, and removing the decoder also reduced prediction accuracy.

  9. Knowl 9 — SCINet representations have lower measured permutation entropy

    empirical result

    The authors measured permutation entropy (PE) for original inputs and SCINet-enhanced representations, using embedding dimension mm and time lag τ=1\tau=1. They report lower PE after SCINet on every listed dataset. Lower PE is interpreted as lower measured sequence complexity and potentially greater predictability, but the paper cautions that it does not guarantee that one time series will always be easier to forecast than another, since accuracy also depends on factors such as available training data, trends, seasonality, and the forecaster.

    Could not parse LaTeX table

  10. Knowl 10 — SCINet is limited by irregular sampling, substantial missingness, and deterministic output

    limitation

    The evaluated SCINet method is designed for regularly sampled time series arranged in chronological order and produces deterministic forecasts. The authors describe it as relatively robust to noise, but warn that sufficiently high missing-data ratios can bias its downsampling-based multiresolution representation and degrade predictions. The even–odd downsampling design may also be difficult to apply to irregularly sampled series. The reported model does not produce probabilistic forecasts. Although it performs competitively on spatial-temporal forecasting without explicitly modeling spatial relations, the authors note that dedicated spatial models could further improve accuracy.

Coverage note — Appendix-level implementation details and additional sensitivity studies for stack count, tree depth, and interaction operators are omitted because they are supporting configuration analyses rather than separate load-bearing contributions.

References

  1. 1.Mohammad Taha Bahadori and Zachary Chase Lipton. Temporal-clustering invariance in irregular healthcare time series. arXiv preprint arXiv:1904.12206, 2019.
  2. 2.Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In ICLR, 2015.
  3. 3.Lei Bai, Lina Yao, Can Li, Xianzhi Wang, and Can Wang. Adaptive graph convolutional recurrent network for traffic forecasting. Advances in Neural Information Processing Systems, 33:17804–17815, 2020.
  4. 4.Shaojie Bai, J Zico Kolter, and Vladlen Koltun. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv preprint arXiv:1803.01271, 2018.
  5. 5.Shaojie Bai, J Zico Kolter, and Vladlen Koltun. Trellis networks for sequence modeling. In ICLR, 2018.
  6. 6.Christoph Bandt and Bernd Pompe. Permutation entropy: a natural complexity measure for time series. Physical review letters, 88(17):174102, 2002.
  7. 7.Anastasia Borovykh, Sander Bohte, and Cornelis W Oosterlee. Conditional time series forecasting with convolutional neural networks. stat, 1050:16, 2017.
  8. 8.George EP Box and Gwilym M Jenkins. Some recent advances in forecasting and control. Journal of the Royal Statistical Society. Series C (Applied Statistics), 17(2):91–109, 1968.
  9. 9.Chao Chen, Karl Petty, Alexander Skabardonis, Pravin Varaiya, and Zhanfeng Jia. Freeway performance measurement system: mining loop detector data. Transportation Research Record, 1748(1):96–102, 2001.
  10. 10.Pierpaolo D’Urso, Livia De Giovanni, and Riccardo Massari. Trimmed fuzzy clustering of financial time series based on dynamic time warping. Annals of operations research, 299(1):1379–1395, 2021.
  11. 11.S. Guo, Youfang Lin, Ning Feng, Chao Song, and Huaiyu Wan. Attention based spatial-temporal graph convolutional networks for traffic flow forecasting. In AAAI, 2019.
  12. 12.Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016.
  13. 13.Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780, 1997.
  14. 14.C. Holt. Forecasting seasonals and trends by exponentially weighted moving averages. International Journal of Forecasting, 20:5–10, 2004.
  15. 15.Rongzhou Huang, Chuyin Huang, Yubao Liu, Genan Dai, and Weiyang Kong. Lsgcn: Long short-term traffic prediction with graph convolutional networks. In IJCAI, pages 2355–2361, 2020.
  16. 16.Yu Huang and Zuntao Fu. Enhanced time series predictability with well-defined structures. Theoretical and Applied Climatology, pages 1–13, 2019.
  17. 17.Rob J Hyndman and Anne B Koehler. Another look at measures of forecast accuracy. International journal of forecasting, 22(4):679–688, 2006.
  18. 18.Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. In ICLR, 2019.
  19. 19.Guokun Lai, Wei-Cheng Chang, Yiming Yang, and Hanxiao Liu. Modeling long- and short-term temporal patterns with deep neural networks. In SIGIR, 2018.
  20. 20.Mengzhang Li and Zhanxing Zhu. Spatial-temporal fusion graph neural networks for traffic flow forecasting. In AAAI, 2021.
  21. 21.Shiyang Li, Xiaoyong Jin, Yao Xuan, Xiyou Zhou, Wenhu Chen, Yu-Xiang Wang, and Xifeng Yan. Enhancing the locality and breaking the memory bottleneck of transformer on time series forecasting. In NeurIPS, 2019.
  22. 22.Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. In ICLR, 2018.
  23. 23.Bryan Lim, Sercan Ö Arık, Nicolas Loeff, and Tomas Pfister. Temporal fusion transformers for interpretable multi-horizon time series forecasting. International Journal of Forecasting, 2021.
  24. 24.Bryan Lim and Stefan Zohren. Time-series forecasting with deep learning: a survey. Philosophical Transactions of the Royal Society A, 379(2194):20200209, 2021.
  25. 25.Shizhan Liu, Hang Yu, Cong Liao, Jianguo Li, Weiyao Lin, Alex X Liu, and Schahram Dustdar. Pyraformer: Low-complexity pyramidal attention for long-range time series modeling and forecasting. In International Conference on Learning Representations, 2021.
  26. 26.Spyros Makridakis, Allan Andersen, Robert Carbone, Robert Fildes, Michele Hibon, Rudolf Lewandowski, Joseph Newton, Emanuel Parzen, and Robert Winkler. The accuracy of extrapolation (time series) methods: Results of a forecasting competition. Journal of forecasting, 1(2):111–153, 1982.
  27. 27.Nam Nguyen and Brian Quanz. Temporal latent auto-encoder: A method for probabilistic multivariate time series forecasting. In AAAI, 2021.
  28. 28.Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. 2016.
  29. 29.Boris Oreshkin, Dmitri Carpo, Nicolas Chapados, and Yoshua Bengio. N-beats: Neural basis expansion analysis for interpretable time series forecasting. In ICLR, 2020.
  30. 30.Frank Pennekamp, Alison C Iles, Joshua Garland, Georgina Brennan, Ulrich Brose, Ursula Gaedke, Ute Jacob, Pavel Kratina, Blake Matthews, Stephan Munch, et al. The intrinsic predictability of ecological time series and its potential to guide forecasting. Ecological Monographs, 89(2):e01359, 2019.
  31. 31.Syama Sundar Rangapuram, Matthias W Seeger, Jan Gasthaus, Lorenzo Stella, Yuyang Wang, and Tim Januschowski. Deep state space models for time series forecasting. NeurIPS, 2018.
  32. 32.David Salinas, Valentin Flunkert, Jan Gasthaus, and Tim Januschowski. Deepar: Probabilistic forecasting with autoregressive recurrent networks. International Journal of Forecasting, 36(3):1181–1191, 2020.
  33. 33.Rajat Sen, Hsiang-Fu Yu, and Inderjit S Dhillon. Think globally, act locally: A deep neural network approach to high-dimensional time series forecasting. volume 32, 2019.
  34. 34.Shun-Yao Shih, Fan-Keng Sun, and Hung yi Lee. Temporal pattern attention for multivariate time series forecasting. Machine Learning, pages 1–21, 2019.
  35. 35.Chao Song, Youfang Lin, Shengnan Guo, and Huaiyu Wan. Spatial-temporal synchronous graph convolutional networks: A new framework for spatial-temporal network data forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 914–921, 2020.
  36. 36.Sean J. Taylor and Benjamin Letham. Forecasting at scale. The American Statistician, 72:37 – 45, 2018.
  37. 37.Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017.
  38. 38.Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. In NeurIPS, 2021.
  39. 39.Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, and C. Zhang. Graph wavenet for deep spatial-temporal graph modeling. In IJCAI, 2019.
  40. 40.Jiehui Xu, Jianmin Wang, Mingsheng Long, et al. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. Advances in Neural Information Processing Systems, 34, 2021.
  41. 41.Ting Yu, Haoteng Yin, and Zhanxing Zhu. Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. In IJCAI, 2018.
  42. 42.Hao-Yi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. Informer: Beyond efficient transformer for long sequence time-series forecasting. In AAAI, 2021.

Citation

MLA
LIU, M., et al. “SCINet: Time Series Modeling and Forecasting with Sample Convolution and Interaction”. Advances in Neural Information Processing Systems, vol. 35, 2022, pp. 5816–28, https://proceedings.neurips.cc/paper_files/paper/2022/file/266983d0949aed78a16fa4782237dea7-Paper-Conference.pdf.
APA
LIU, M., Zeng, A., Chen, M., Xu, Z., LAI, Q., Ma, L., & Xu, Q. (2022). SCINet: Time Series Modeling and Forecasting with Sample Convolution and Interaction. Advances in Neural Information Processing Systems, 35, 5816–5828. https://proceedings.neurips.cc/paper_files/paper/2022/file/266983d0949aed78a16fa4782237dea7-Paper-Conference.pdf
Chicago
LIU, M., A. Zeng, M. Chen, et al. 2022. “SCINet: Time Series Modeling and Forecasting with Sample Convolution and Interaction”. Advances in Neural Information Processing Systems 35: 5816–28. https://proceedings.neurips.cc/paper_files/paper/2022/file/266983d0949aed78a16fa4782237dea7-Paper-Conference.pdf.
Harvard
LIU, M. et al. (2022) “SCINet: Time Series Modeling and Forecasting with Sample Convolution and Interaction”, Advances in Neural Information Processing Systems. Curran Associates, Inc., pp. 5816–5828. Available at: https://proceedings.neurips.cc/paper_files/paper/2022/file/266983d0949aed78a16fa4782237dea7-Paper-Conference.pdf.
Vancouver
1. LIU M, Zeng A, Chen M, Xu Z, LAI Q, Ma L, Xu Q (2022) SCINet: Time Series Modeling and Forecasting with Sample Convolution and Interaction. In: Advances in Neural Information Processing Systems. Curran Associates, Inc., pp 5816–5828

BibTeX

@inproceedings{liu2022scinet,
  title = {SCINet: Time Series Modeling and Forecasting with Sample Convolution and Interaction},
  author = {LIU, Minhao and Zeng, Ailing and Chen, Muxi and Xu, Zhijian and LAI, Qiuxia and Ma, Lingna and Xu, Qiang},
  year = {2022},
  booktitle = {Advances in Neural Information Processing Systems},
  publisher = {Curran Associates, Inc.},
  volume = {35},
  pages = {5816-5828},
  url = {https://proceedings.neurips.cc/paper_files/paper/2022/file/266983d0949aed78a16fa4782237dea7-Paper-Conference.pdf}
}
Metadata:DOI registry

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF
License: Authors