Exploring the Training Performance of Deep Neural Networks on Embedded Many-core Processors

• Item type: Symposium(1)
• 公開日: 2023-12-20
• タイトル: Exploring the Training Performance of Deep Neural Networks on Embedded Many-core Processors
• タイトル / 言語: en
• タイトル: Exploring the Training Performance of Deep Neural Networks on Embedded Many-core Processors
• 言語: eng
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_5794
• 資源タイプ: conference paper
• 著者所属: Saitama University
• 著者所属: Saitama University
• 著者所属(英): en
• 著者所属(英): Saitama University
• 著者所属(英): en
• 著者所属(英): Saitama University
• 著者名: Masahiro, Hasumi; Takuya, Azumi
• 著者名(英): Masahiro, Hasumi; Takuya, Azumi
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Vehicular Internet of Things (IoT) is playing an important role in the burgeoning field of autonomous vehicle development. One of the key technologies underpinning this development is deep learning (DL), which is instrumental for tasks like precise object detection. However, the training data used for these models often harbors sensitive vehicle-related information, creating privacy concerns if consolidated on a central server, due to the potential for information leakage. Moreover, adhering to privacy measures can ramp up the power consumption of IoT devices onboard the vehicle. In response to these challenges, our study explores the potential of using many-core processors for training DL models, offering a low-power alternative to traditional GPUs. Many-core processors, furnished with multiple computational cores, provide an opportunity for low-power operations. This study serves as a foundational assessment aimed at enabling Federated Learning (FL) on edge devices in vehicular IoT systems. Specifically, we focus on evaluating the speed of model training on many-core processors as a critical preliminary step toward FL. Our evaluation includes an assessment of training speed and scalability on clustered many-core processors. This demonstrates that these processors are viable for on-edge DL model training in vehicular IoT, setting the groundwork for future applications that require both data privacy and reduced power consumption. Thus, our findings contribute to a new avenue for balancing computational performance and energy efficiency in the increasingly complex landscape of vehicular IoT.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Vehicular Internet of Things (IoT) is playing an important role in the burgeoning field of autonomous vehicle development. One of the key technologies underpinning this development is deep learning (DL), which is instrumental for tasks like precise object detection. However, the training data used for these models often harbors sensitive vehicle-related information, creating privacy concerns if consolidated on a central server, due to the potential for information leakage. Moreover, adhering to privacy measures can ramp up the power consumption of IoT devices onboard the vehicle. In response to these challenges, our study explores the potential of using many-core processors for training DL models, offering a low-power alternative to traditional GPUs. Many-core processors, furnished with multiple computational cores, provide an opportunity for low-power operations. This study serves as a foundational assessment aimed at enabling Federated Learning (FL) on edge devices in vehicular IoT systems. Specifically, we focus on evaluating the speed of model training on many-core processors as a critical preliminary step toward FL. Our evaluation includes an assessment of training speed and scalability on clustered many-core processors. This demonstrates that these processors are viable for on-edge DL model training in vehicular IoT, setting the groundwork for future applications that require both data privacy and reduced power consumption. Thus, our findings contribute to a new avenue for balancing computational performance and energy efficiency in the increasingly complex landscape of vehicular IoT.
• 書誌情報: Proceedings of Asia Pacific Conference on Robot IoT System Development and Platform; 巻 2023, p. 14-21, 発行日 2023-12-20
• 出版者 / 言語: ja
• 出版者: 情報処理学会