Promises of Meta-Learning for Device-Free Human Sensing
Farid Ghareh Mohammadi, M. Hadi Amini · 2019
In this study, we explore potential opportunities for leveraging meta-learning algorithms to enable device-free sensing. We refer to this solution as "meta-sensing", which is mainly learning to sense by discovering the available information rather than deploying additional sensors. We specifically are interested in application of zero-shot learning, a specific algorithm that required no a priori information to learn. This class of methods aim at learning to learn, i.e., meta-sensing does not only learn from the available data, but also learns how to learn over time without requiring extra sensing inputs. Meta-sensing learns and predicts through data transformation with respect to the test data. It executes the process of mapping data and updates information; learns how to transform input data to new feature vector space; and generates a new data cluster. Meta-learner, an agent who updates the learning model, plays a pivotal role in meta-sensing. Once new data are received, meta-learner updates its bias information using transformed data. The meta-learner learns based on ontology first rather than a priori information or training data set.