A Transfer Learning-Based New User Recognition for Minimizing Retraining Time in Edge Deep Learning

Manually adding new users to user recognition systems can ensure high accuracy, but it is a cumbersome process. Recent research is focused on methods for automatically detecting and adding new users. Researchers consider high-performance servers essential because they utilize massive deep learning models to ensure accuracy. Moreover, obtaining a substantial amount of new user data requires both the edge node and the server to consume a significant amount of time during data exchange. To address this, researchers have explored various methods, such as conducting online learning on the edge node. However, there are many constraints due to the resource limitations of the edge node. Thus, this paper suggests a system where learning takes place on the server, and all other functions are carried out on the edge node. The proposed system suggests the use of the transfer learning method to minimize the time required for adding new users with a high similarity foot pressure dataset. Through this method, the required amount of new user data for retraining was reduced by 25 %. Additionally, a system was developed to dynamically update the deep learning model received from the server in real-time on the edge node. As a result, using a model trained on the existing 10 users as a basis, retraining 10 new users with 140 training data each achieved a recognition performance of 86 % for 20 users. Additionally, it was shown that by reducing the training data, it is possible to add a new user within 7.7 minutes, which is an 80 % decrease in time.