Search

Large-scale benchmark datasets are crucial in advancing research within the computer science communities. They enable the development of more sophisticated AI models and serve as "golden" benchmarks for evaluating their performance. Thus, ensuring the quality of these datasets is of utmost importance for academic research and the progress of AI systems. For the emerging vision-language tasks, some datasets have been created and frequently used, such as Flickr30k, COCO, and NoCaps, which typically contain a large number of images paired with their ground-truth textual descriptions. In this paper, an automatic method is proposed to assess the quality of large-scale benchmark datasets designed for vision-language tasks. In particular, a new cross-modal matching model is developed, which is capable of automatically scoring the textual descriptions of visual images. Subsequently, this model is employed to evaluate the quality of vision-language datasets by automatically assigning a score to each 'ground-truth' description for every image picture. With a good agreement between manual and automated scoring results on the datasets, our findings reveal significant disparities in the quality of the ground-truth descriptions included in the benchmark datasets. Even more surprising, it is evident that a small portion of the descriptions are unsuitable for serving as reliable ground-truth references. These discoveries emphasize the need for careful utilization of these publicly accessible benchmark databases. [ABSTRACT FROM AUTHOR]

Building precise machine learning and deep learning models has traditionally required a combination of mathematical skills and hands-on experience to meticulously adjust hyperparameters that significantly impact the learning process. As datasets continue to expand across various engineering domains, researchers increasingly turn to machine learning methods to uncover hidden insights that may elude classic regression techniques. This surge in adoption raises concerns about the adequacy of resultant meta-models and the interpretation of findings. In response to these challenges, automated machine learning (AutoML) emerges as a promising solution, aiming to construct machine learning models with minimal intervention or guidance from human experts. This paper benchmarks AutoML solutions by providing an overview of their principles and applying them to predict the most important mechanical properties of different concrete datasets, i.e., compressive strength. Nine datasets from various concrete types, sample sizes, and features are utilized, with a detailed discussion on the benchmark dataset from high-performance concrete, applying best practices to the other eight datasets. For each case, the importance of hyperparameter tuning is discussed, alongside the ensemble and stacking models. Tree-based models are employed for each dataset to develop SHAP plots, interpret results, and understand the contribution of each component in the mix design to the overall strength of the concrete. This paper further explores three unique aspects of benchmarking AutoML in material science: (1) "reliability" by contrasting the benchmark dataset's error metric with literature collected over the past 20 years, (2) "uncertainty" by quantifying the variability in the mean and standard deviation of the error metric from different datasets and its correlation with the sample-to-feature ratio, and (3) "dilemma" by discussing the shortcomings of AutoML in specific concrete datasets. • The first comprehensive application of AutoML is presented in the concrete regression model. • Nine different concrete datasets are compared in terms of compressive strength. • AutoML's reliability, and dilemmas are evaluated in predicting concrete strength. • A benchmark problem is discussed to quantify the AutoML uncertainty in material science. • Explainable machine learning is used to identify the key contributing factors in f' c. [ABSTRACT FROM AUTHOR]