We consider the problem of machine unlearning to erase a target dataset, which causes an unwanted behavior, from the trained model when the training dataset is not given. Previous works have assumed that the target dataset indicates all the training data imposing the unwanted behavior. However, it is often infeasible to obtain such a complete indication. We hence address a practical scenario of unlearning provided a few samples of target data, so-called few-shot unlearning. To this end, we devise a straightforward framework, including a new model inversion technique to retrieve the training data from the model, followed by filtering out samples similar to the target samples and then relearning. We demonstrate that our method using only a subset of target data can outperform the state-of-the-art methods with a full indication of target data. * equal contribution Preprint. Under review. CE(f(1)Given w o and D e ⊂ D (without D), the standard unlearning task [23,12] aims at approximating w u trained with the same objective in equation 1 but D r := D \ D e instead of D, i.e.,This is typically formulated to unlearn an unwanted behavior of w o imposed by D e . For instance, letting D e be the set of all the mislabeled data in training, the goal of unlearning is to correct the
The steel I-girder bridge can collapse if girder is seriously damaged by disasters. In this paper, a numerical study on the redundancy of the steel multi-girder bridges after damage was performed. The redundancy evaluation method used in NCHRP Report 406 was implemented to evaluate the redundancy of the girder intact and damaged bridges. The considered damage cases include reduction of material strength and lower flange damage. The damaged multi-girder bridges were evaluated considering both material and geometric nonlinearity and the redundancy was evaluated for each case. According to the analytical results of this study, steel I-girder bridge has sufficient redundancy after reduced strength in material and the number of lower flange.
Recently, researchers are conducting studies to improve the mechanical and chemical properties of cementitious composites mixed with nanomaterials. Defects may occur inside nano-cementitious composites due to nanomaterial agglomeration in the manufacturing process. These defects can degrade the mechanical performance of the nano-cementitious composite. This study performs ultrasonic non-destructive and compressive strength tests according to the size of defects in nano-cementitious composites. Multi-walled carbon nanotubes (MWCNTs) were used for the nanomaterial, and internal defects of various sizes were considered in the center of the specimens. Ultrasonic pulse velocity was measured according to the defect size until 30 curing days, after which the compressive strength was measured. The ultrasonic pulse velocity of the nano-cementitious composites decreased by up to 9.6% in relation to that of the specimens without defects as the defect size increased, and the compressive strength decreased by up to 35.7%. This study’s findings revealed a correlation between ultrasonic pulse velocity and compressive strength according to defect size. Future ultrasonic non-destructive tests will allow for the prediction of mechanical performance and the detection of defects within nano-cementitious composites.
Abstract:Ultra High Performance Concrete (UHPC) is a superior structural material with high strength and durability. The main objective of this study is to investigate behavior of the lap-spliced reinforced joints in UHPC. The lap-spliced joint type has simple details but it is expected to be highly effective method of joints in UHPC structure because of the high bond stress of UHPC. The static test was performed to verify the effect of the joint in the UHPC bridge deck slab. The major parameter considered in experimental plan was lap-spliced length and the strength of UHPC. Test results show that the minimum lap-spliced length which behaved similar to the continued reinforcement test specimen in strength and ductility was 150 mm (5.91 in.).
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.