摘要：

将深度学习部署到边缘侧存在训练数据的域偏移、长尾分布、计算资源有限的问题, 因此需要应用领域自适应方法进行在线重训缓解域偏移, 重训时利用长尾削减技术缓解长尾分布问题, 且需要考虑计算开销。然而现有长尾削减技术大多计算开销较大或无法与领域自适应方法有效结合。为此, 提出一种边缘侧结合领域自适应和长尾分布削减技术的算法EdgeTailor。EdgeTailor将边缘合成少数类过采样技术和类平衡损失作为长尾削减优化策略, 通过对连续无监督自适应过程进行优化, 并引入缓冲区解决在线学习时尾部类数据量过少的问题, 使其能够在进行在线连续领域自适应的同时缓解在线学习时数据的长尾问题。实验结果表明, 在两组存在域偏移的长尾数据集, 以5种深度神经网络作为模型骨架构建的边缘侧领域自适应任务中, EdgeTailor相比基线在目标域上平均Top-1准确率提升了约8.10%;在计算开销方面, EdgeTailor的每秒浮点运算次数(FLOPs)和参数量均保持在相对较低的水平, FLOPs相比基线中效果较好的数据合成方法减小了大约29.84%。EdgeTailor具有边缘侧高性能和低开销的优点, 有效缓解领域自适应中的长尾视觉识别问题。

关键词: 边缘智能, 深度学习, 长尾, 领域自适应, 在线学习

Abstract:

The deployment of deep learning models at the edge is hindered by challenges such as domain offset, long-tail distribution, and limited computing resources in the training data. Therefore, domain adaptation methods must be applied for online retraining to alleviate the domain offset, and long-tail reduction techniques must be applied during retraining to alleviate the long-tail problem while considering computational costs. However, most existing long-tail reduction techniques have high computational costs or cannot be effectively combined with domain adaptation methods. To address these issued, this paper proposes EdgeTailor, a long-tail optimization method specifically designed for edge-side domain adaptation. EdgeTailor optimizes the continuous unsupervised adaptive process by using synthetic minority class oversampling techniques and class-balanced loss as strategies for tail truncation. Consequently, a buffer is introduced to address the issue of insufficient data for tail classes during online learning, allowing it to mitigate the long-tail problem while conducting online continuous domain adaptation. Experimental results demonstrate the effectiveness of EdgeTailor in edge domain adaptation tasks involving two long-tail datasets with domain shift. Using five deep neural networks as the model backbone, EdgeTailor improves average Top-1 accuracy by approximately 8.10% compared with the baseline in the target domain. In terms of computational cost, EdgeTailor maintains a low level of Floating Point Operations Per Second (FLOPs) and parameter count, reducing FLOPs by approximately 29.84% compared with the data synthesis method, with better performance than the baseline. Overall, EdgeTailor achieves high performance and low cost in addressing both domain adaptation and long-tail visual recognition challenges in edge deployment.

Key words: edge intelligence, deep learning, long-tail, domain adaptation, online learning