@inproceedings{2e82fb4e63ea46bdb01bc31669c0b902,
title = "XST: A Crossbar Column-wise Sparse Training for Efficient Continual Learning",
abstract = "Leveraging the ReRAM crossbar-based In-Memory-Computing (IMC) to accelerate single task DNN inference has been widely studied. However, using the ReRAM crossbar for continual learning has not been explored yet. In this work, we propose XST, a novel crossbar column-wise sparse training framework for continual learning. XST significantly reduces the training cost and saves inference energy. More importantly, it is friendly to existing crossbar-based convolution engine with almost no hardware overhead. Compared with the state-of-the-art CPG method, the experiments show that XST's accuracy achieves 4.95 % higher accuracy. Furthermore, XST demonstrates 5.59 × training speedup and 1.5 × inference energy-saving.",
keywords = "Continual Learning, In-Memory-Computing, Sparse Learning",
author = "Fan Zhang and Li Yang and Jian Meng and Seo, {Jae Sun} and Yu Cao and Deliang Fan",
note = "Funding Information: Another major hardware disadvantage induced by CPG learning is the element-wise reprogramming and resetting. We quantified the energy consumption based on the writing voltage, writing pulses, and conductance level changes [13], [14]. As shown in Figure 4, the element-wise reprogramming and resetting of the CPG learning [1] lead to inconsistent and massive energy consumption during the entire continual learning phase (T2 to T20). The proposed XST resets the weights by turning off the corresponding ReRAM columns. As a result, the hardware resetting cost of the XST becomes zero. During the continual learning process, XST replenishes the weights in a group-wise manner, corresponding to the consistent reprogramming cost of the ReRAM columns. Compared to the CPG learning, the XST-trained model achieves 2.5× reprogramming energy reduction along sub-tasks. V. CONCLUSION In summary, we propose XST, a hardware-friendly cross-bar column-wise sparse learning method to efficiently deploy Fig. 4: Energy consumption of the reprogramming and resetting continual learning to ReRAM crossbar based neural network accelerator with the consideration of hardware cost. Comparing with CPG, our XST shows 4.95% accuracy improvement, ∼5.59X training speedup, 1.5× inference energy saving, and 1.8X× re-programming energy saving on CIFAR-100 dataset with VGG16-BN(1.5×) model. ACKNOWLEDGMENT This work is supported in part by the National Science Foundation under Grant No.2003749, No.1931871, No. 2144751 Publisher Copyright: {\textcopyright} 2022 EDAA.; 2022 Design, Automation and Test in Europe Conference and Exhibition, DATE 2022 ; Conference date: 14-03-2022 Through 23-03-2022",
year = "2022",
doi = "10.23919/DATE54114.2022.9774660",
language = "English (US)",
series = "Proceedings of the 2022 Design, Automation and Test in Europe Conference and Exhibition, DATE 2022",
publisher = "Institute of Electrical and Electronics Engineers Inc.",
pages = "48--51",
editor = "Cristiana Bolchini and Ingrid Verbauwhede and Ioana Vatajelu",
booktitle = "Proceedings of the 2022 Design, Automation and Test in Europe Conference and Exhibition, DATE 2022",
}