引導式測驗 · 硬 等級
VC 維度與 PAC 學習測驗
測試破碎、VC 維度、不可知的 PAC 樣本複雜性,以及為什麼最壞情況的保證可能比觀察到的模型性能更寬鬆。
問題 1 的 8
For three points to be shattered by a binary hypothesis class, what must the class be able to realize?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- For three points to be shattered by a binary hypothesis class, what must the class be able to realize?
- The class can correctly classify all n points on unseen test data
- The class contains exactly n hypotheses
- The points are linearly separable by exactly one hypothesis
- For every one of the 2^n possible labelings of those points, some hypothesis in the class reproduces it exactly
- For straight-line binary classifiers in a two-dimensional plane, how many points can be shattered at most?
- 1
- 3
- 2
- 4
- In the agnostic PAC setting described in the guide, how does a standard worst-case sample bound depend on VC dimension and epsilon?
- Sample size needed scales roughly with VC-dimension divided by epsilon squared
- Sample size needed decreases as VC dimension increases
- Sample size needed is independent of VC dimension
- Sample size needed depends only on the number of features, never on VC dimension
- Why is a single-threshold classifier on the real line said to have VC dimension 1?
- It has one parameter, t, so its VC dimension equals its parameter count
- It can only classify positive examples
- It can shatter exactly one point but not every arrangement of two points
- It has an infinite VC dimension since t can be any real number
- Is VC dimension the same thing as the number of parameters in a model, according to the guide?
- Yes, they are always identical
- Yes, but only for linear models
- No, VC dimension only applies to neural networks
- No; some infinite-parameter classes have finite VC dimension and some low-parameter models can have high VC dimension
- What theoretical result explains why the growth function becomes only polynomial in n once n exceeds the VC dimension?
- Sauer's lemma
- 中心極限定理
- 貝葉斯定理
- 大數定律
- Why do classical VC-dimension-based PAC bounds struggle to explain deep neural network generalization?
- Neural networks have zero VC dimension
- PAC theory does not apply to any parametric models
- Neural networks' VC dimension is often far larger than practical training set sizes, yet they still generalize well
- Neural networks cannot be represented as hypothesis classes
- What kind of guarantee does PAC learning provide about a learned hypothesis?
- That it will have exactly zero error
- That with high probability its error is within some epsilon of the best hypothesis in the class
- That it will always outperform every other hypothesis class
- That it requires no training data at all