Викторина, связанная с руководством · Средний Уровень
Викторина по теореме об отсутствии бесплатных обедов
Описывается усреднение теоремы в конечном пространстве задач, индуктивное смещение и то, почему проверка для конкретной задачи все еще может выявить сильные алгоритмы для реальных данных.
Вопрос 1 из 8
Under the classic finite-domain No Free Lunch result, what happens when search algorithms are averaged uniformly over the complete set of objective functions?
Предварительный просмотр каждого вопроса8 вопросы
Ответы и пояснения появляются по мере прохождения теста.
- Under the classic finite-domain No Free Lunch result, what happens when search algorithms are averaged uniformly over the complete set of objective functions?
- Their average performance is equal under the theorem’s assumptions
- An algorithm using gradients always wins
- The random algorithm always wins
- The algorithm with the most parameters wins
- Why does a convolutional network often outperform other models on natural images despite NFL?
- Images are randomly structured so any model works equally well
- NFL only applies to text data
- Its built-in assumption of spatial locality matches how real images are structured
- Convolutional networks have no inductive bias
- Who formalized the No Free Lunch theorems referenced in this guide, and when?
- Джеффри Хинтон в 2006 году
- David Wolpert and William Macready in 1997
- Алан Тьюринг в 1950 году
- Ян ЛеКун в 1998 году
- What does the No Free Lunch result imply about a benchmark leaderboard?
- It proves one algorithm is best on every task
- It makes cross-validation unnecessary
- It means measured benchmark differences are errors
- It supports conclusions about the tested task distribution, not universal superiority
- When does an algorithm’s inductive bias help its performance, according to the guide?
- When the bias matches regularities in the task problems it encounters
- When the algorithm has the largest number of parameters
- When every possible objective function is equally likely in production
- When the training set contains no repeated examples
- What does 'inductive bias,' as described in this guide, refer to?
- A dataset labeling error
- A hardware limitation on model size
- The average number of features in a dataset
- An algorithm's built-in preference for certain kinds of patterns or hypotheses
- How does the technical section describe the NFL proof's treatment of the objective function?
- As a uniformly random draw from all possible functions over a fixed domain
- As a fixed, known function chosen by the researcher
- As always being a linear function
- As irrelevant to the proof
- Per the guide, what should NFL practically motivate a practitioner to do when picking a model for a hospital readmission-risk task?
- Assume the winner of a published benchmark elsewhere will transfer directly
- Pick the newest published architecture regardless of data
- Test several algorithms against their own specific dataset rather than assume a universal winner
- Ignore cross-validation since NFL makes it pointless