Many existing neural architecture search (NAS) solutions rely on downstream training for architecture evaluation, which takes enormous computations. Considering that these computations bring a large carbon footprint, this paper aims to explore a green (namely environmental-friendly) NAS solution that evaluates architectures without training. Intuitively, gradients, induced by the architecture itself, directly decide the convergence and generalization results. It motivates us to propose the gradient kernel hypothesis: Gradients can be used as a coarse-grained proxy of downstream training to evaluate random-initialized networks. To support the hypothesis, we conduct a theoretical analysis and find a practical gradient kernel that has good correlations with training loss and validation performance. According to this hypothesis, we propose a new kernel based architecture search approach KNAS. Experiments show that KNAS achieves competitive results with orders of magnitude faster than "train-then-test" paradigms on image classification tasks. Furthermore, the extremely low search cost enables its wide applications. The searched network also outperforms strong baseline RoBERTA-large on two text classification tasks. Codes are available at \url{https://github.com/Jingjing-NLP/KNAS} .
The deterioration mechanisms of sulfate attack on concrete under sustained loading and wet-dry cycling were investigated based on micro and macroobservations. The mass fraction of sodium sulfate solution was 6.9%. Three loading levels of 20, 40, and 60% of ultimate flexural load were considered, and the load was mechanically applied to the specimens under four-point bending. Microobservations included the analysis of the chemical products formed using thermal analysis and the determination of the sulfate-ion content profile using the modified barium sulfate gravimetric method (chemical titration). Macroobservations primarily included visual observations and flexural strength changes. Test results showed that under alternate action of wet-dry cycling, concretes are attacked by expansive products such as ettringite and gypsum during the wetting cycle, and crystallization damage, induced by evaporation, is superposed during the drying cycle. Results also showed that the tensile stresses can increase diffusivity by initiating or developing microcracks; the compressive stresses are small compared with the concrete compressive strength, so any effect on ion transport properties is not obvious. Under simultaneous sulfate attack and flexural loading, deterioration is aggravated as the loading level increases, and this is characterized on the macroscale by the increased strength degradation. This research can provide some references for the assessment method of concrete structures under combined sulfate attack and loading action.
Detecting out-of-distribution (OOD) instances is significant for the safe deployment of NLP models. Among recent textual OOD detection works based on pretrained language models (PLMs), distance-based methods have shown superior performance. However, they estimate sample distance scores in the last-layer CLS embedding space and thus do not make full use of linguistic information underlying in PLMs. To address the issue, we propose to boost OOD detection by deriving more holistic sentence embeddings. On the basis of the observations that token averaging and layer combination contribute to improving OOD detection, we propose a simple embedding approach named Avg-Avg, which averages all token representations from each intermediate layer as the sentence embedding and significantly surpasses the state-of-the-art on a comprehensive suite of benchmarks by a 9.33% FAR95 margin. Furthermore, our analysis demonstrates that it indeed helps preserve general linguistic knowledge in fine-tuned PLMs and substantially benefits detecting background shifts. The simple yet effective embedding method can be applied to fine-tuned PLMs with negligible extra costs, providing a free gain in OOD detection. Our code is available at https://github.com/lancopku/Avg-Avg.
This article presents the results of a comparative experimental study to investigate the effects of using three different fire protection measures to improve the fire endurance of timber assembly. The three fire protection measures were fire-retardant intumescent coating, gypsum plasterboard and filling the timber assembly void with mineral wool. The fire-resistance period obtained from the fire endurance tests was based on the integrity and the insulation criteria. Compared to the reference timber assembly without any fire protection (fire-resistance time = 37 min), the increases in fire resistance using the three different fire protection measures were 6 min (16%), 37 min (100%) and 142 min (384%) for using intumescent coating (specimen F2), 12-mm-thick gypsum plasterboard (specimen F4) and infill mineral wool (F3), respectively. The specific intumescent coating used in the test failed to expand. Therefore, this specimen (F2) behaved very similarly with the control specimen (F1) without any fire protection. Attaching an additional layer of gypsum plasterboard to the timber assembly on the fire-exposed side improved the fire-resistance rating by about 30 min, which is higher than that obtained from using the current design guidance such as Eurocode EN 1995-1-2. Among these three fire protection methods, filling the void between the top and bottom timber boards gave the best result because the mineral wool not only provided insulation but also stopped direct flame attack of the timber board on the unexposed side.
Abstract Beta diversity is the key and central issue of theoretical and applied ecological questions, including those regarding spatial variation along gradients and conservation biodiversity. In contrast to the alpha diversity of stream fish assemblages, little concern has been reached regarding the factor affecting of spatial variation in beta diversity along longitudinal gradients, which may be modified by scale, indices of beta diversity and pairwise methods. To understand the spatial patterns and processes in the species composition of stream fish assemblages along the river continuum, we need to disentangle the influence of different beta diversity indices and pairwise methods on the longitudinal patterns of beta diversity. In this study, we chose four beta diversity indices (Jaccard, Sørensen, Simpson and additive partitioning) and two pairwise methods (mean beta diversity between the focal assemblage and all other assemblages; beta diversity between focal and immediate neighbouring assemblages downstream) to quantify the spatial pattern of fish beta diversity along the river continuum. We found that (1) although the four diversity indices differed in emphasis, they showed the same pattern of spatial variation in the longitudinal gradients of the river; (2) the same diversity index using different pairwise methods calculation may result in completely different spatial patterns; (3) beta diversity indices and pairwise may not have a significant effect on the ecological application of beta diversity, at least in community assembly. In conclusion, the results of our study suggest that at least in the longitudinal gradient of the river there may not be much difference in quantifying the spatial variation of beta diversity; however, pairwise methods between sampling sites need to be chosen according to specific research objectives.
Video paragraph captioning (VPC) involves generating detailed narratives for long videos, utilizing supportive modalities such as speech and event boundaries. However, the existing models are constrained by the assumption of constant availability of a single auxiliary modality, which is impractical given the diversity and unpredictable nature of real-world scenarios. To this end, we propose a Missing-Resistant framework MR-VPC that effectively harnesses all available auxiliary inputs and maintains resilience even in the absence of certain modalities. Under this framework, we propose the Multimodal VPC (MVPC) architecture integrating video, speech, and event boundary inputs in a unified manner to process various auxiliary inputs. Moreover, to fortify the model against incomplete data, we introduce DropAM, a data augmentation strategy that randomly omits auxiliary inputs, paired with DistillAM, a regularization target that distills knowledge from teacher models trained on modality-complete data, enabling efficient learning in modality-deficient environments. Through exhaustive experimentation on YouCook2 and ActivityNet Captions, MR-VPC has proven to deliver superior performance on modality-complete and modality-missing test data. This work highlights the significance of developing resilient VPC models and paves the way for more adaptive, robust multimodal video understanding.
The past 10 years have witnessed the rapid growth of global mobile cellular traffic demands due to the popularity of mobile devices. While accurate traffic prediction becomes extremely important for stable and high-quality Internet service, the performance of existing methods is still poor due to three challenges: complicated temporal variations including burstiness and long periods, multi-variant impact factors such as the point of interest and day of the week, and potential spatial dependencies introduced by the movement of population. While existing traditional methods fail in characterizing these features, especially the latter two, deep learning models with powerful representation ability give us a chance to consider these from a new perspective. In this article, we propose Deep Traffic Predictor (DeepTP), a deep-learning-based end-toend model, which forecasts traffic demands from spatial-dependent and long-period cellular traffic. DeepTP consists of two components: a general feature extractor for modeling spatial dependencies and encoding the external information, and a sequential module for modeling complicated temporal variations. In the general feature extractor, we introduce a correlation selection mechanism for a spatial modeling and embedding mechanism to encode external information. Moreover, we apply a seq2seq model with attention mechanism to build the sequential model. Extensive experiments based on large-scale mobile cellular traffic data demonstrate that our model outperforms the state-of-the-art traffic prediction models by more than 12.31 percent.
'/%3e%3cuse xlink:href='%23a' transform='rotate(23.036 .093 25.536)'/%3e%3cuse xlink:href='%23a' transform='rotate(45.87 1.273 16.18)'/%3e%3cuse xlink:href='%23a' transform='rotate(69.945 .996 12.078)'/%3e%3cuse xlink:href='%23a' transform='rotate(20.66 -19.689 31.932)'/%3e%3c/svg%3e)