適合增量學習的高效邊緣模型部署研究

Efficient Model Deployment on the Edge for Incremental Learning

費浩杰(Haojie Fei)

指導教授 ： 楊佳玲

國立臺灣大學/電機資訊學院/資訊工程學研究所/碩士(2022年)

https://doi.org/10.6342/NTU202104561

摘要

近年來，深度學習在圖像分類等任務上擁有越來越不錯的表現。此外，隨著普及計算和IoT的發展，越來越多的深度學習模型需要部署在邊緣設備上。在將深度學習的模型放到邊緣設備上時，我們要對深度學習的模型進行壓縮來滿足設備的要求。另一方面，我們希望部署後的深度學習模型也仍能夠持續不斷的學習，也即當有新的類別資料出現時，我們的模型能夠學習辨識新類別資料的能力。增量學習可以幫助我們擴展模型的能力同時避免遺忘舊的知識。因此，我們提出NPIL，一種適合增量學習的，利用了濾波器剪枝模型壓縮方法的高效深度學習模型生成方法。我們的方法采用了訓練前剪枝的方式，同時考慮了舊類別知識的權重和新類別知識的權重，來產生適合的剪枝策略。在完成模型剪枝後，我們再使用新的小模型來進行增量學習。我們采用了AlexNet，VGG，ResNet三個模型和CUB，StanfordCars 兩個資料集進行實驗，相比於基礎方法和另一個簡單的利用小模型來進行增量學習的方法，我們的方法可以維持較好的準確率並同時保持不錯的速度。在時間上，對於三個模型， NPIL 比基礎方法快2.2到2.5倍。在準確率上，對於 CUB資料集，NPIL 比簡單方法最多要好3.9%，而對於CAR資料集， NPIL要比簡單方法最多好3.0%。

關鍵字

增量學習 ； 深度學習 ； 模型剪枝 ； 模型壓縮 ； 邊緣運算

並列摘要

In recent years, deep learning has been performing well on tasks such as image classification. In addition, with the development of pervasive computing and IoT, more and more deep learning models need to be deployed on edge devices. When putting deep learning models on edge devices, we have to compress the deep learning models to meet the requirements of the devices. On the other hand, we hope that the deployed deep learning models can learn incrementally, which means our models can learn to recognize new classes of data when new classes appears. Incremental learning can help us extend the capability of our model while avoiding forgetting old knowledge. Therefore, we propose an efficient deep learning model generation method for incremental learning that utilizes a filter pruning model compression method. The proposed approach NPIL employs a pre-training new-class-aware pruning approach that takes into account both the weight of the old class knowledge and the weight of the new class knowledge to generate a suitable pruning strategy. After the model pruning is completed, we then use the new pruned model for incremental learning. We use AlexNet, VGG, ResNet models and CUB, StanfordCars datasets for our experiments. Compared with the baseline method and another naive incremental learning method using pruned models, the proposed NPIL can maintain better accuracy and better performance. For time comparison, NPIL is 2.2x - 2.5x faster than baseline for three models. For accuracy comparison on CUB dataset, NPIL's accuracy is at most 3.9 percent better than the naive solution for three models. While for CAR dataset, NPIL can achieve at most 3.0 percent better accuracy than naive solution.

並列關鍵字

Incremental Learning ； Deep Learning ； Model Pruning ； Model Compression ； Edge Computing

參考文獻

[1] S.-A. Rebuffi, A. Kolesnikov, G. Sperl, and C. H. Lampert, “icarl: Incremental classifier and representation learning,” in Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 2001-2010, 2017.

[2] D. Isele and A. Cosgun, “Selective experience replay for lifelong learning,” in arXiv, 2018.

[3] H. Shin, J. K. Lee, J. Kim, and J. Kim, “Continual learning with deep generative replay,” in arXiv, 2017.

[4] D. Lopez-Paz and M. Ranzato, “Gradient episodic memory for continual learning,” Advances in Neural Information Processing Systems (NeurIPS), pages 6467-6476, 2017.

[5] S.-W. Lee, J.-H. Kim, J. Jun, J.-W. Ha, and B.-T. Zhang, “Overcoming catastrophic forgetting by incremental moment matching,” in arXiv, 2017.