We investigate the performance of DNNs when trained on class-incremental visual problems consisting of initial training, followed by retraining with added visual classes. Catastrophic forgetting (CF) behavior is measured using a new evaluation procedure that aims at an application-oriented view of incremental learning. In particular, it imposes that model selection must be performed on the initial dataset alone, as well as demanding that retraining control be performed only using the retraining dataset, as initial dataset is usually too large to be kept. Experiments are conducted on class-incremental problems derived from MNIST, using a variety of different DNN models, some of them recently proposed to avoid catastrophic forgetting. When comparing our new evaluation procedure to previous approaches for assessing CF, we find their findings are completely negated, and that none of the tested methods can avoid CF in all experiments. This stresses the importance of a realistic empirical measurement procedure for catastrophic forgetting, and the need for further research in incremental learning for DNNs.

No abstract

The CSIR Built-Environment, in conjunction with the University of Pretoria and the Cement and Concrete Institute of South Africa, developed a low cost option for the upgrading of unsurfaced (gravel) roads. The proposed solution is the placing of a thin layer of normal concrete reinforced with 5.6mm diameter steel with a mesh grid size of 200mm. This thin layer is placed on top of the existing unsurfaced road with minimal preparation to the existing road surface using labourintensive construction methods.Through full-scale trials this type of upgrading proved to be adequate for low-volume traffic applications (e.g. residential streets) as well as for higher-volume applications (e.g. bus routes). During the trials test sections were subjected to a total of over 700,000 ESALs over a period of 5 years without showing any deterioration.In order to determine the structural capacity of this type of overlay full-scale Heavy Vehicle Simulator tests were conducted. This paper summarizes the initial results from the accelerated pavement testing (APT) tests and is aimed at building confidence in the use of thin-layer CRCP, with cognizance being taken of the pavement structure, support conditions, construction, climate and traffic.

No abstract

Welding process causes notch effects and residual stresses. Accordingly, the risk of brittle fracture grows. In comparison with currently used steels, old mild steels have a lower toughness. The weldability of mild steel is limited and only feasible under specific conditions. If for structural reasons welding is preferred instead of the use of bolts, the metallurgical characteristics of the steels have to be considered. These include the concentrations of impurities in the zones of segregation and the tendency to embrittlement by nitrogen-induced ageing.In the corresponding paper, experimental and analytical studies of the weldability of old mild steels are presented. Extensive material analyses to determine the mechanical and the technological properties of the material will be an essential part of the investigations. Particularly, the increased impurities of phosphorus, sulfur, nitrogen and also oxygen in the segregation zones as well as the distinctive non-metallic inclusions complicate to produce load-bearing butt welds. Welding materials currently available are not designed for welding such materials. Nevertheless, in the context of a research project welding tests on old mild steels has been carried out using current welding electrodes. Supported by involved industrial partners, a stick electrode adapted to the characteristics of old steels will be developed.