Utilizing the Hive Mind – How to Manage Knowledge in Fully Distributed Environments

Abstract

By 2020, the Internet of Things will consist of 26 Billion connected devices. All these devices will be collecting an innumerable amount of raw observations, for example, GPS positions or communication patterns. In order to benefit from this enormous amount of information, machine learning algorithms are used to derive knowledge from the gathered observations. This benefit can be increased further, if the devices are enabled to collaborate by sharing gathered knowledge. In a massively distributed environment, this is not an easy task, as the knowledge on each device can be very heterogeneous and based on a different amount of observations in diverse contexts. In this paper, we propose two strategies to route a query for specific knowledge to a device that can answer it with high confidence. To that end, we developed a confidence metric that takes the number and variance of the observations of a device into account. Our routing strategies are based on local routing tables that can either be learned from previous queries over time or actively maintained by interchanging knowledge models. We evaluated both routing strategies on real world and synthetic data. Our evaluations show that the knowledge retrieved by the presented approaches is up to 96.7% as accurate as the global optimum.

Knowledge Is at the Edge! How to Search in Distributed Machine Learning Models

“…In order to retrieve knowledge this notion has to be extended by some sort of confidence metric that can take the quality of available knowledge (information) into account. Such a confidence metric needs to express the expertise of a node, reflecting for example that it holds a lot of similar information [5] or can do reliable prediction. In our previous work [5] we have tackled this issue for knowledge modeled as N-Dimensional point-clouds.…”

“…Such a confidence metric needs to express the expertise of a node, reflecting for example that it holds a lot of similar information [5] or can do reliable prediction. In our previous work [5] we have tackled this issue for knowledge modeled as N-Dimensional point-clouds. We proposed a point-cluster-based confidence metric that took the variance and number of points in each cluster as an indicator of quality into account.…”