NAS-Bench-360

NAS-Bench-360是由多伦多大学的研究团队开发的一个包含10个不同任务的数据集，用于评估神经架构搜索（NAS）方法在超越传统研究领域的多样性任务上的性能。该数据集涵盖了多种应用领域、数据集大小、问题维度和学习目标，特别设计以适应现代卷积神经网络（CNN）搜索方法。数据集的核心在于其任务导向的特性，旨在评估不同NAS方法和搜索空间在广泛领域的表现。此外，数据集还提供了两个任务的预计算性能结果，包含15,625种架构，以加速和降低NAS研究的计算成本。NAS-Bench-360的目的是推动NAS方法在未充分探索或资源有限的领域中的应用，并促进科学发现。

NAS-Bench-360 is a dataset developed by a research team at the University of Toronto, which includes 10 distinct tasks for evaluating the performance of neural architecture search (NAS) methods on diverse tasks beyond traditional research domains. This dataset covers a wide range of application fields, dataset sizes, problem dimensions and learning objectives, and is specially designed to adapt to modern convolutional neural network (CNN) search methods. The core of the dataset lies in its task-oriented characteristic, which aims to evaluate the performance of different NAS methods and search spaces across a broad spectrum of fields. In addition, the dataset provides precomputed performance results for two tasks, containing 15,625 architectures, to accelerate NAS research and reduce its computational cost. The purpose of NAS-Bench-360 is to promote the application of NAS methods in under-explored or resource-constrained fields and facilitate scientific discovery.

搜集汇总

背景与挑战

背景概述

NAS-Bench-360是一个由多伦多大学开发的神经架构搜索（NAS）基准数据集，包含10个多样性任务，覆盖多种应用领域和问题维度，特别设计用于评估NAS方法在传统研究之外的广泛表现。其核心特点是任务导向，并提供部分预计算性能结果（15,625种架构）以降低研究计算成本，旨在推动NAS在未充分探索或资源有限领域的应用和科学发现。