[2504.10561] Self-Controlled Dynamic Expansion Model for Continual Learning

View a PDF file from the paper entitled The Dynamic Expansion Form that controls myself

PDF HTML (experimental) view

a summary:Continuing learning (CL) summarizes an advanced training model as previous data samples remain accessible while acquiring new tasks. Many investigations have taken advantage of the pre -vision transformer (VIT) to enhance the effectiveness of the model in continuous learning. However, these methods usually use the unique spine, which is inappropriately adapting to new tasks, especially when engaging with various data fields, due to a large number of inactive parameters. This paper deals with this restriction by introducing an innovative, self -controlled dynamic expansion model (SCDEM), which organizes many of the outstanding spine that was trained before training to provide various and engraved various representations. Specifically, by employing multiple bone structure as a common unit, SCDEM generates a dynamic expert with a minimum of parameters to accommodate a new task. A new cooperative improvement mechanism (COM) is presented to improve the multiple spine in a consisting manner by harnessing the signs of prediction of historical experts, thus facilitating learning new tasks without erasing the pre -acquired knowledge. In addition, the approach to the consistency of the new features (FDC) is proposed to align the semantic similarities between the previously learned representations through a mechanism based on the optimal transfer distance, which effectively reduces the effects of negative knowledge transfer. Moreover, to alleviate the inaugural challenges in the system, this paper provides an attention mechanism of a new dynamic feature (DLWFAM) to determine the severity of the punishment on each training -to -tract representation layer. A wide range of experiments was conducted to assess the effectiveness of the proposed methodology, with experimental results confirming that the approach reaches the latest performance.