AddNIST, Language, MultNIST, CIFARTile, Gutenberg, Isabella, GeoClassing, Chesseract

本研究引入了八个新的数据集，用于神经架构搜索（NAS）挑战，旨在解决NAS领域过度依赖少数基准数据集的问题。这些数据集包括AddNIST、Language等，涵盖了从图像处理到语言分析的多样化任务。每个数据集都经过精心设计，以模拟真实世界问题的复杂性，如AddNIST通过将MNIST图像的通道相加来增加难度。这些数据集的创建不仅推动了NAS技术的发展，也鼓励研究者开发能够适应未知数据集的模型，从而提高NAS方法的泛化能力。

This study introduces eight novel datasets for neural architecture search (NAS) challenges, aiming to address the issue that the NAS field overly relies on a small number of benchmark datasets. These datasets include AddNIST, Language, and others, covering diverse tasks ranging from image processing to language analysis. Each dataset is meticulously designed to simulate the complexity of real-world problems; for example, AddNIST increases task difficulty by summing the channels of MNIST images. The creation of these datasets not only promotes the advancement of NAS technologies but also encourages researchers to develop models capable of adapting to unseen datasets, thereby enhancing the generalization capability of NAS approaches.