贝叶斯深度学习测验

当您参加测验时，就会出现答案和解释。

1. 贝叶斯神经网络对待权重的方式与标准神经网络有何不同？
   • 它将每个权重视为概率分布而不是单个固定值
   • 它使用整数权重而不是小数
   • 它完全消除了重量
   • 它在所有层之间共享一个权重
2. What can a Bayesian approach provide that a single fixed-weight network does not directly represent?
   • A guarantee that every confidence score is calibrated
   • A distribution of predictions that reflects uncertainty about model weights
   • A proof that the model will detect every unfamiliar input
   • A way to remove all noise from the data
3. How can epistemic uncertainty change as relevant, representative training data are added?
   • It is guaranteed to reach zero with any added data
   • It is identical to label noise
   • It may decrease under the chosen model and data assumptions
   • It always increases linearly
4. 为什么需要变分推理或 MCMC 等近似方法？
   • 因为神经网络没有损失函数
   • 因为GPU不能做加法
   • 因为权重必须是整数
   • 因为数百万个权重的精确后验在计算上是难以处理的
5. 蒙特卡罗 dropout 是如何估计不确定性的？
   • 在测试时保持 dropout 处于活动状态，并利用输出的分布多次运行网络
   • 关闭辍学并延长训练时间
   • 永久移除所有神经元
   • 用随机噪声代替损失
6. 变分推理将后验估计重新定义为什么样的问题？
   • 一个排序问题
   • 最小化 KL 散度的优化问题（通过 ELBO）
   • 没有目标的纯随机抽样问题
   • 硬件设计问题
7. How does Bayes’ rule relate a prior, likelihood, and posterior?
   • The posterior is the prior plus the likelihood, without normalization
   • It is the likelihood divided by the prior without normalization
   • The posterior is proportional to the prior times the likelihood
   • It is the fixed point estimate of the most likely weight vector
8. How does MCMC approximate a Bayesian neural-network posterior?
   • By training a single point estimate with dropout disabled
   • By ranking model outputs with a fixed rule
   • By minimizing an ELBO with a simpler variational family
   • By drawing samples through a Markov chain whose stationary distribution targets the posterior