Machine learning acceleration method with multidimensional data embedding and coded distributed computing for internet of things

Abstract

As the number of connected devices grows with the proliferation of Internet of Things (IoT), a massive amount of data is being rapidly generated. Recently, machine learning has been introduced into IoT for analysis of the bigdata. To apply machine learning technologies to the multidimensional time-series data generated by IoT, feature extraction is essential to convert the time-series data into a form that facilitates learning. Accordingly, a feature extraction method is required to reduce the time complexity in time series processing and enhance the performance of machine learning technologies. Meanwhile, the need for distributed data processing in IoT devices has recently emerged due to latency or data protection issues. To this end, distributed computing using the computing resources of IoT devices has come to the fore for machine learning inference or training. However, IoT devices are often characterized by tight resource constraints, which limit their capabilities to manage cumbersome machine learning models. Moreover, in large-scale distributed computing, voluntary participation of IoT devices is difficult to expect, and slow nodes that delay the computing inevitably occur due to their heterogeneity. To manage these issues and accelerate the process in machine learning for IoT, this dissertation proposes a feature extraction framework for IoT data with multidimensional data embedding and a distributed computing mechanism with coding and incentive mechanisms for large-scale machine learning training. The proposed feature extraction framework was experimentally shown to significantly reduce the time complexity in time series analysis with machine learning while improving the analysis performance. In addition, the proposed distributed computing mechanisms were also shown to effectively reduce the training time while guaranteeing the profit of the IoT devices that participate in large-scale distributed computing. The proposed feature extraction framework and distributed computing mechanism can be respectively utilized to support accelerated machine learning inference and training for IoT.