The high performance of mobile devices has enabled deep learning to be extended to also exploit its strengths on such devices. However, because their computing power is not yet sufficient to perform on-device training, a pre-trained model is usually downloaded to mobile devices, and only inference is performed on them. This situation leads to the problem that accuracy may be degraded if the characteristics of the data for training and those for inference are sufficiently different. In general, fine-tuning allows a pre-trained model to adapt to a given data set, but it has also been perceived as difficult on mobile devices. In this paper, we introduce our on-going effort to improve the quality of mobile deep learning by enabling fine-tuning on mobile devices. In order to reduce its cost to a level that can be operated on mobile devices, a light-weight fine-tuning method is proposed, and its cost is further reduced by using distributing computing on mobile devices. The proposed technique has been applied to LetsPic-DL, a group photoware application under development in our research group. It required only 24 seconds to fine-tune a pre-trained MobileNet with 100 photos on five Galaxy S8 units, resulting in an excellent image retrieval accuracy reflected a 27–35% improvement.

移动设备的高性能使深度学习得以扩展，从而也可以利用其在此类设备上的优势。但是，由于它们的计算能力还不足以执行设备上的训练，因此通常会将预训练的模型下载到移动设备，并且仅对它们执行推断。这种情况导致以下问题：如果用于训练的数据和用于推断的数据的特性足够不同，则准确性可能下降。通常，微调允许预训练的模型适应给定的数据集，但是在移动设备上也被认为很难。在本文中，我们介绍了通过在移动设备上进行微调来提高移动深度学习质量的持续努力。为了将其成本降低到可以在移动设备上操作的水平，提出了一种轻量级的微调方法，并且通过在移动设备上使用分布式计算来进一步降低其成本。所提出的技术已应用于LetsPic-DL，这是我们研究小组正在开发的小组照相软件应用程序。仅需24秒，就可以在五个Galaxy S8设备上微调带有100张照片的预训练MobileNet，从而获得了出色的图像检索精度，反映出27-35％的改进。