A summary of Sustainable and Trustworthy Edge Machine Learning

Ivona Brandic; IEEE Computing Edge, January 2023 DOI [0]

Table of Contents

• A summary of Sustainable and Trustworthy Edge Machine Learning
  • Table of Contents
  • Summary
    • Data Intensive Applications
    • Hyper Heterogeneous Hardware
    • Challenges In Terms Of Sustainability And Trustworthiness
    • Trustworthiness In Rucon: Fault Tolerance
    • Sustainability In Rucon: Model Distribution For Edge Machine Learning
    • Sustainability And Trustworthiness In Rucon: Data Quality Management
  • Citations

Summary

Understanding the trustworthiness of the system traditionally involved creating complex mathematical abstractions of the system. The author’s of this article propose an orthogonal, data driven, and probabilistic model called Rucon (Runtime Control on Multi-Clouds). Rucon is able to scale accordingly and model complex, data driven systems.

IoT devices are revolutionizing many sectors by providing near real time data feeds. These feeds, to be of use, need to be analyzed in real time. Because of this, there is a move towards understanding and modeling this “tactile internet”.

One of the problems with the tactile internet is that it generates Extreme Data, which is a step above Big Data. The challenges of Extreme Data is that it requires near exascale computing resources to process this data. Additionally, this data is not generated from a single source or even geographic location. Therefore, there is a problem in modeling the flow data within systems that utilize both Extreme Data and the tactile internet.

Data Intensive Applications

Applications on the tactile internet are not solely based on cloud computing resources like traditional ML applications.

Hyper Heterogeneous Hardware

Edge devices simply collect data; they do not compute any data on device. Therefore, they are dependent upon nearby devices to process the data to perform ML analysis. This is known as Edge Machine Learning (EML).

There is a move to equipping centers of IoT activity with 5G connectivity to send data to onsite micro data centers that consist of a few Raspberry Pis.

Challenges In Terms Of Sustainability And Trustworthiness

In traditional ML apps running on the cloud, the solution to sustainability is to shut down VMs that have low resource utilization. However, tactile internet apps receive data from many sensors continuously.

To model both the sustainability and trustworthiness challenges of both ML for the Edge and at the Edge, the authors created the Rucon architecture. Rucon was developed with hyper heterogeneous systems in mind, and is capable of dealing with the geographic distribution, intermittent connectivity, and low availability of processing nodes within a tactile internet application. It introduces:

• A novel fault tolerance and trustworthiness method based on Dynamic Bayesian Networks to address ML for the Edge.
• A sustainable model management approach for geographically distributed ML to address ML at the Edge.
• A sustainable and trustworthy method for data quality management for failure prone IoT devices for both ML at and for the Edge.

Trustworthiness In Rucon: Fault Tolerance

Edge device systems fail more often than large traditional systems. To address this, it is best practice to implement many geographically close edge devices and perform spatial-temporal rollover to the next closest device. In other words, when one device fails, continue processing on the next device. However, the problem with this system is that geographically close nodes often fail together. To address this, Rucon using Dynamic Bayesian Networks to predict when a node will fail as well as what other nodes will be taken offline by a single node failure. This information can be used to preemptively address nodes prior to failure. They were able to decrease the redundancy loss by 50% compared to their baseline.

Sustainability In Rucon: Model Distribution For Edge Machine Learning

ML models are developed and trained on large cloud computers. Reduced versions of these models are then deployed to edge devices. However, in a geographically distributed system, these models can go in an out of sync with one another. This means that some models may lose accuracy depending upon the data that they process over time. An example would be that edge devices in urban environments would be trained to recognize pedestrians better than devices deployed in rural environments. Thus, to keep all devices in sync, a traditional system would constantly update nodes with the latest model. However, on edge devices were connectivity is limited, this would incur a significant messenger cost. Therefore, Rucon uses a reinforcement learning algorithm for update scheduling and synchronization of edge nodes. After training on a dataset of smart vehicles, Rucon was able to save 90% of model updates while maintaining a suitable level of model synchronization.

Sustainability And Trustworthiness In Rucon: Data Quality Management

Data quality can become inconsistent with a geographically distributed system. To address this, the Rucon has a number of data recovery techniques to recover gaps in datasets. These include the auto regressive integrated moving average, exponential smoothing method, and data clustering to identify outliers.

Citations