Broadcasting and Media Magazine (방송과미디어)

The Korean Institute of Broadcast and Media Engineers (한국방송∙미디어공학회)
권호 표지

MPEG Compression of Neural Network (NNC) 국제표준 기술 동향

  • Published : 2023.01.30
Download PDF
⟨ Previous Next ⟩

Abstract

인공신경망 모델이 다양한 분야에서 뛰어난 성능을 보이고 있지만, 동시에 모델의 복잡도도 크게 증가하였다. 따라서, 모바일 같은 저전력 디바이스에 인공신경망 모델이 실시간으로 추론/배포되기 위해서는 모델의 가중치 파라미터의 수 혹은 메모리 소모량을 줄이는 경량화 기술이 필수적이다. 이에 MPEG에서는 인공신경망 모델을 다양한 프레임워크에서 상호 운용 가능하고 파라미터를 압축 표현하는 NNC (Compression of Neural Networks) 표준화를 진행 중에 있다. 본고에서는 NNC 표준의 개요와 가중치 파라미터를 압축하는 압축 기술, 그리고 HLS (High-Level Syntax)들을 소개하고자 한다.

Keywords

Acknowledgement

이 연구는 2023년도 정부(과학기술정보통신부)의 재원으로 정보통신기획평가원의 지원을 받아 수행된 연구임 (No.2021-0-00802, 속성을 유지하는 지능적 미디어 화면비 변환 기술 개발)

References

  1. S. Han, et al, "Deep Compression: Compressing Deep Neural Net-works with pruning, trained quantization and Huffman coding," In Proc. CVPR, Jun. 2015.
  2. https://www.tensorflow. org/
  3. https://pyotrch.org/
  4. Neural Network Exchange Format (The Khronos NNEF Working Group), [Available at Online] https://www.khronos.org/registry/NNEF/specs/1.0/nnef-1.0.3.pef
  5. Open Neural Network Exchange, [Available at Online] https://github.com/onnx/onnx/blob/ master/onnx/onnx.proto
  6. W. Bailer, et al, "Text of ISO/IEC FDIS 15938-17 Compression of Neural Networks for Multimedia Content Description and Analysis," ISO/IEC JTC1/SC29/WG4, N20331, Jun. 2021.
  7. W.Bailer, et al, "Draft Call for Proposals on Incremental Compression of Neural Networks for multimedia content description and analysis," ISO/IEC JTC1/SC29/WG11, N19515, Jul. 2020.
  8. W. Bailer, et al, "Use cases and requirements for Compressed Representation of Neural Networks," ISO/IEC JTC1/SC29/WG11, N17924, Oct. 2019.
  9. W. Bailer, et al, "Evaluation Framework for Compression of neural networks for multimedia content description and analysis," ISO/IEC JTC1/SC29/WG11, N18575, Jul. 2019.
  10. S. Niknam, et al, "Federated learning for wireless communications: Motivation, opportunities, and challenges," IEEE Communications Magazine, 58(6): 46-51, 2020. https://doi.org/10.1109/MCOM.001.1900461
  11. H. Moon, et al, "Test Data for Incremental NNR: Federated Learning for Medical Applications (UC 14A)," ISO/IEC JTC1/SC29/WG11 m55054, Oct. 2020.
  12. W. Bailer, et al, "Test Model 6 of Compression of Neural Networks for Multimedia Content Description and Analysis," ISO/IEC JTC1/SC29/WG11, N19765, Oct. 2020.
  13. C. Aytekin, et al, "Response to the Call for Proposals on Neural Network Compression: Training Highly Compressible Neural Networks," ISO/IEC JTC1/SC29/WG11, m47379, Mar. 2019.
  14. W. Jiang, et al, "[NNR] CE1 result: micro_structured_pruning," ISO/IEC JTC1/SC29/WG11, m55022, Oct. 2020.
  15. S. Lin, et al, "Holistic CNN Compression via Low-Rank Decomposition with Knowledge Transfer," IEEE transaction on pattern analysis and machine intelligence, 41(2): 2889-2905, 2019. https://doi.org/10.1109/TPAMI.2018.2873305
  16. M. Jaderberg, et al, "Speeding up Convolutional Neural Networks with Low Rank Expansions," In Proc. CVPR, Jun. 2014.
  17. H. Moon, et al, "Response to the Call for Proposals on Neural Network Compression: Quantization and Low-Rank Approximation," ISO/IEC JTC1/SC29/WG11, m47704, Mar. 2019.
  18. W. Jiang, et al, "NNR non-CE1 related: Data-dependent transformation for highly unified Neural Networks," ISO/IEC JTC1/SC29/WG11, m52631, Jun. 2020.
  19. H. Schwarz, T. Nguyen, D. Marpe and T. Wiegand, "CE7: Transform Coefficient Coding and Dependent Quantization (Tests 7.1.2, 7.2.1)", JVET-K0071, 2018.
  20. P. Hasse, et al, "[NNR] CE2-related: Dependent scalar quantization for neural network parameter approximation," ISO/IEC JTC1/SC29/WG11, m52358, Jun. 2020.
  21. P. Hasse, et al, "[NNR]: HLS adaptation for integer codebook representation," ISO/IEC JTC1/SC29/WG11, m54937, Jun. 2020.
  22. S. Wiedemann, et al, "DeepCABAC: Context-adaptive Binary Arithmetic Coding for Deep Neural Network Compression," International Conference on Machine Learning (ICML), May. 2019.
  23. H. Kirchhoffer, et al, "Overview of the Neural Network Compression and Representation (NNR) Standard," IEEE transcation on circuits and systems for video technology, 32(5): 3203-3216, 2022. https://doi.org/10.1109/TCSVT.2021.3095970
  24. W. Bailer, et al, "Common Test Conditions for Incremental Neural Network Compression," ISO/IEC JTC1/SC29/WG4, N0123, Aug. 2021.