Abstract:

Today, many services use content recommendation systems for marketing purposes. This powerful tool is also used in educational platforms to filter educational content. Thus, educational resources are selected in relation to the learner's objectives, while adapting them to the learner's preferred learning style (Messika, 2020). This selection is possible through the use of semantic technologies that offer an ontological language to structure a set of concepts representing a field of information (Héon, 2019) and a logical description language for the standardized description of resources. Such an approach allows the processing of resources by programs (Apoki, 2021) through an interconnection of these without human intervention (duChâteau, 2020) while improving the possibility of inference. There are already semantic active learning systems (e.g. SASA) capable of enriching and personalizing the learner's experience by exploiting a reasoner using the calculation of first-order predicates and ontologies modeling the entities participating in the learning process (Szilagyi, 2012). The current proposal is to use peripheral artificial intelligence with a semantic system in order to realize a personal learning agent. This tutor will use knowledge monitoring by storing all the data issued by the learner in an LRS (Learning Record Store) in different forms of models (e.g. Overlay, Perturbation, Constraint Based, Fuzzy Modeling) in order to ensure consulted resources monitoring and achievement of targeted educational objectives (Alkhatlan, 2019). This information will then be synthesized in order to have a clear follow-up of the learner's evolution (Franzoni, 2020), follow-up which will also be used by the tutor to train and perfect his model. Thanks to the Internet of Things, it is possible to use different connected objects as learning aids. The use of edge computing offers several advantages, including avoiding the transmission of sensitive learner information in clouds while allowing the use of multiple devices (Dustdar, 2019). This method, used in cloud computing to process data as close as possible to the emission source, avoids the transmission of a large amount of data because only those that are relevant and pre-processed are sent to the cloud (Ismael, 2018). This technique applied to artificial intelligence will work in two stages: it begins with a local learning where each device adjusts its learning model, followed by a global aggregation where the main server defines the weights of the new model and updates it on the various connected objects (Li, 2019). Data is not transferred between devices, only models are transferred: time and bandwidth are saved, and the private aspect of the data is protected (Hosseinalipour, 2020). However, certain limits should be identified: a model based on biased data can result in learning biases linked to data or societal biases (Mélot, 2021), certain sensitive information can be found by reverse engineering model parameters (Tu, 2020) or the risk of misunderstanding of this technology by a society where computing remains low-level office automation (O'Neil, 2016).

Keywords:

Personalized Learning, Intelligent Tutoring Systems (ITS), Peripheral Artificial Intelligence, Edge Computing.