引导式测验 · 中等 等级

神经架构搜索测验

查看 NAS 搜索空间、算法、早期搜索成本、DARTS 松弛、权重共享和评估注意事项。检查假设和范围。

相关引导路径神经架构搜索
问题 1 的 8

神经架构搜索自动化什么？

预览每个问题8 问题

当您参加测验时，就会出现答案和解释。

  1. 神经架构搜索自动化什么？
    • 神经网络结构的设计
    • 标记训练数据
    • 冷却 GPU 服务器
    • 为模型编写文档
  2. 哪三个组件定义了 NAS 方法？
    • 数据集、优化器和 dropout
    • 搜索空间、搜索策略和性能估计
    • 编码器、解码器和注意力
    • 训练、验证和测试集
  3. What accelerator scale did the original reinforcement-learning NAS paper report for its large search?
    • One GPU for every NAS method
    • A fixed cost of exactly 1,000 GPU-days for every search
    • An experiment using 800 GPUs in that study
    • No accelerator use during architecture search
  4. DARTS 使用什么关键技巧来提高搜索效率？
    • 它在单个示例上进行训练
    • 它删除所有隐藏层
    • 它仅使用随机搜索
    • 它将离散的架构选择放松为可通过梯度优化的连续架构
  5. 权重共享 NAS 中的“超级网”是什么？
    • 一个包含所有候选架构的超参数化网络，其权重是共享的
    • 具有数十亿参数的网络
    • 仅在超级计算机上运行的网络
    • 模型的备份副本
  6. What did the HURRICANE hardware-aware NAS study evaluate?
    • One architecture that is guaranteed optimal on every device
    • Tailored architectures for three hardware types under matched latency constraints
    • A language model trained without any hardware measurements
    • A method that removes latency from architecture evaluation
  7. What is one limitation of weight-sharing NAS supernets?
    • They remove all architecture choices
    • They guarantee the globally best architecture
    • The shared weights can bias architecture rankings relative to training each candidate independently
    • They do not use neural networks
  8. What does a NAS evaluation need to report for a fair comparison?
    • Only the best test score
    • The GPU brand alone
    • The model name without settings
    • Search space, compute budget, validation protocol, and final retraining setup