[2507.21455] Boost Self-Supervised Dataset Distillation via Parameterization, Predefined Augmentation, and Approximation

PDF display of the paper entitled enhancing the distillation of the self-supervision data set by identifying the parameter, pre-specified enlargement, and approximation, by Xing-Ving Yu and 2 other authors

PDF HTML (experimental) view

a summary:Although the larger data sets are decisive to train large deep models, the rapid growth of the size of the data set has brought a major challenge in terms of large training costs, which leads to exorbitant mathematical expenses. Data setting becomes a popular technique in recent times to reduce the size of the data set by learning a very compressed set of representative models, as the trained model on these models should perfectly have a similar performance with regard to the trainer who has been trained with the full data set. While most of the works focus on distillation of the data group on the data -subject groups, we instead aim to distort the images and its self -trained representations in a bishop. This procedure, which has been named as a self -controlled database, has effectively extracts rich information from real data collections, which leads to distilled groups with a circular of cross architecture. In particular, in order to maintain the main characteristics of the original data group with more honesty and integration, many new technologies are suggested: 1) We offer an innovative parameter on images and representations through distinct low -dimensional rules, where the primary selection of the teacher is experimental to play a decisive role; 2) We deal with the instability caused by the random to increase the data-which is a major component of self-controlled learning, but it is reduced in the previous work to distill the set of self-supervision data-by using pre-defined reinforcements; 3) We take advantage of a lightweight network for the modeling of the communications between the reinforced views of the same image, which leads to more distillation pairs. Intensive experiences conducted on different data groups confirm the validity of our approach in terms of distillation efficiency, generalization of structure through the structure, and the transfer of learning performance.