機器學習 (machine learning, ML) 是人工智慧 (artificial intelligence, AI)的一種，透過大量的資料跟經驗來從中反覆學習並且找到分類規則或是訓練模型，在之後輸入新的資料時可以透過規則或模型進行預測。現今機器學習在各個領域都在快速進展，針對機器學習的加速器也被大量研究，這些加速器通常處在長時間的運算，加速老化效應的發生，老化會造成計算延遲增加，重則會產生功能錯誤，而在機器學習中的老化會使精確度下降，在實際應用層面如自駕車以及醫療模型中出現精確度下降是不可接受的，所以處理該問題是當務之急。
機器學習需要大量乘加運算，乘法所需時間更是加法無法比擬的。現今的乘法器多由壓縮器組合而成，壓縮器的近似運算也是眾多研究的目標，因為機器學習的特性，可以犧牲一部份的精度換取時間跟功耗。而老化所造成的延遲也可透過犧牲精度來減少運算時間補償，但是目前的研究缺少針對老化而做出的優化，在本篇論文中，我們提出一個有效的可動態重新配置的4:2近似壓縮器，可以在還未老化時精確運算，在老化時透過近似運算減少運算時間用以補償老化增加的延遲。實驗結果表明，我們的方法可以保證老化10年的近似運算精確度不變。

Machine Learning is a subset of Artificial Intelligence (AI) that involves learning from large amounts of data and experience to identify classification rules or train models. When new data is inputted, these models can make predictions based on the learned rules. Currently, machine learning is advancing rapidly across various fields, and accelerators for machine learning are being extensively researched. These accelerators often operate for long periods, accelerating aging effects. Aging can lead to increased computational latency and can also lead to functional errors in severe cases. In machine learning, aging can cause a decrease in accuracy, which is unacceptable in practical applications such as autonomous vehicles and medical models. Therefore, addressing this issue is urgent.

Machine learning requires numerous multiply-accumulate operations, with multiplication being significantly more time-consuming than addition. Modern multipliers are often composed of compressors, and approximate computing for compressors is a major research focus. Given the characteristics of machine learning, some precision can be sacrificed for improved time and power efficiency. Aging-induced delays can also be mitigated by sacrificing precision to compensate for increased computation time. However, current research lacks optimization specifically for aging. In this paper, we propose an effective dynamically reconfigurable 4:2 approximate compressor that performs accurate computations before aging and uses approximate computations to reduce computation time and compensate for aging-induced delays. Experimental results show that our method ensures that the accuracy of approximate computations remains unchanged after 10 years of aging.

[1] Y. Lecun, L. Bottou, Y. Bengio, P. Haffner, “Gradient-based learning applied to document recognition”, in Proc. of Proceedings of the IEEE, vol. 86, issue. 11, pp. 2278-2324, Nov. 1998
[2] A. Krizhevsky, I. Sutskever and G. E. Hinton, “Imagenet classification with deep convolutional neural networks”, Proc. Adv. Neural Inf. Process. Syst., pp. 1097-1105, 2012.
[3] K. He, et al., “Deep Residual Learning for Image Recognition”, in CVPR, 2016.
[4] K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition”, in ICLR, 2015.
[5] C. S. Wallace, “A Suggestion for a Fast Multiplier”, IEEE Transactions on Electronic Computers ( Volume: EC-13, Issue: 1, February 1964)
[6] Luigi Dadda, “Some schemes for fast serial input multipliers”, 1983 IEEE 6th Symposium on Computer Arithmetic (ARITH)
[7] Antonio Giuseppe Maria Strollo, Ettore Napoli, Davide De Caro, Nicola Petra, Gennaro Di Meo, “Comparison and Extension of Approximate 4-2 Compressors for Low-Power Approximate Multipliers”, IEEE Transactions on Circuits and Systems I: Regular Papers ( Volume: 67, Issue: 9, September 2020)
[8] Hang Xiao, Haobo Xu, Xiaoming Chen, Yujie Wang, Yinhe Han, “Fast and High-Accuracy Approximate MAC Unit Design for CNN Computing”, IEEE Embedded Systems Letters ( Volume: 14, Issue: 3, September 2022)
[9] Haoran Pei, Xilin Yi, Hang Zhou, Yajuan He, “Design of Ultra-Low Power Consumption Approximate 4–2 Compressors Based on the Compensation Characteristic”, IEEE Transactions on Circuits and Systems II: Express Briefs ( Volume: 68, Issue: 1, January 2021)
[10] U. Anil Kumar, Sumit K. Chatterjee, Syed Ershad Ahmed, “Low-Power Compressor-Based Approximate Multipliers With Error Correcting Module”, IEEE Embedded Systems Letters ( Volume: 14, Issue: 2, June 2022)
[11] Dieter K. Schroder, Jeff A. Babcock, “Negative bias temperature instability: Road to cross in deep submicron silicon semiconductor manufacturing”, Journal of applied Physics, vol. 94, pp. 1–18, July 2003.
[12] K. Rott, H. Reisinger, S. Aresu, C. Schlünder, K. Kölpin, W. Gustin, T. Grasser, “New insights on the PBTI phenomena in SiON pMOSFETs”, Microelectron. Rel., vol. 52, nos. 9–10, pp. 1891–1894
[13] T. H. Ning, P. W. Cook, R. H. Dennard, C. M. Osburn, S. E.Schuster, and H. N. Yu, “ 1 um MOSFET VLSI technology: Part 5 Hot-electron design constraints”, IEEE J. Solid-Stare Circuits, vol.SC-14, pp. 268-275, April, 1979.
[14] B. Eitan and D. Frohman-Bentchkowsky, “Hot-electron injection into the oxide in n-channel MOS devices”, IEEE Trans. Electron Devices, vol. ED-28, pp. 328-340. 1981.
[15] Heesu Kim, Jongho Kim, Hussam Amrouch, Jörg Henkel, Andreas Gerstlauer, Kiyoung Choi, Hanmin Park, “Aging Compensation With Dynamic Computation Approximation”, IEEE Transactions on Circuits and Systems I: Regular Papers ( Volume: 67, Issue: 4, April 2020)
[16] Sami Salamin, Georgios Zervakis, Ourania Spantidi, Iraklis Anagnostopoulos, Jörg Henkel, Hussam Amrouch, “Reliability-Aware Quantization for Anti-Aging NPUs”, 2021 Design, Automation & Test in Europe Conference & Exhibition (DATE)
[17] D. Ernst, Nam Sung Kim, S. Das; S. Pant, R. Rao, Toan Pham, C. Ziesler, D. Blaauw, T. Austin, K. Flautner, T. Mudge, “Razor: a low-power pipeline based on circuit-level timing speculation”, Proceedings. 36th Annual IEEE/ACM International Symposium on Microarchitecture, 2003. MICRO-36.
[18] D. Ernst, S. Das, S. Lee, D. Blaauw, T. Austin, T. Mudge, Nam Sung Kim, K. Flautner, “Razor: circuit-level correction of timing errors for low-power operation”, IEEE Micro ( Volume: 24, Issue: 6, Nov.-Dec. 2004)