IncludeHelp_logo
Skip Navigation Links
×

Trending Technologies MCQs

Home » Trending Technologies MCQs

Few-Shot Learning (FSL) and Zero-Shot Learning (ZSL) MCQs (Multiple-Choice Questions)

Practice Few-Shot Learning (FSL) and Zero-Shot Learning (ZSL) MCQs to test your understanding of modern machine learning techniques that enable models to recognize new tasks and categories with limited or no task-specific training examples. These multiple-choice questions cover learning approaches, model architectures, prompt-based methods, applications, advantages, and limitations. They are useful for students, AI developers, machine learning engineers, and candidates preparing for technical interviews. The collection includes foundational and practical questions to strengthen your understanding of FSL and ZSL.

Few-Shot Learning (FSL) and Zero-Shot Learning (ZSL) MCQs

These Few-Shot Learning and Zero-Shot Learning multiple-choice questions cover important concepts such as few-shot classification, zero-shot classification, meta-learning, transfer learning, support sets, query sets, semantic embeddings, large language models, prompting, generalized zero-shot learning, and evaluation metrics. The questions explore how models learn from limited examples, generalize to unseen classes, and use prior knowledge to solve unfamiliar tasks through conceptual, technical, and scenario-based questions.

These FSL and ZSL MCQs help learners understand the differences between learning with a few labeled examples and recognizing categories without labeled examples from those categories during task-specific training. Each question includes an answer and explanation.

List of Few-Shot Learning (FSL) and Zero-Shot Learning (ZSL) MCQs

Explore the following 50 MCQs covering the principles, techniques, architectures, applications, evaluation methods, and practical challenges of Few-Shot Learning and Zero-Shot Learning.

1. What is Few-Shot Learning (FSL)?

  1. A learning approach that requires millions of labeled examples for every task
  2. A learning approach that learns to perform a task using only a small number of labeled examples
  3. A technique that works exclusively with unlabeled data
  4. A method that prevents models from learning new classes

Answer: B) A learning approach that learns to perform a task using only a small number of labeled examples

Explanation:

Few-Shot Learning enables a model to adapt to a new task or recognize new categories using a limited number of labeled examples. It is useful when collecting large labeled datasets is expensive or impractical.

2. What is Zero-Shot Learning (ZSL)?

  1. Learning a new category using exactly one labeled example
  2. Training a model without using any data of any kind
  3. Recognizing or performing tasks for which no task-specific labeled examples are provided for the target category
  4. Memorizing all possible classes before deployment

Answer: C) Recognizing or performing tasks for which no task-specific labeled examples are provided for the target category

Explanation:

Zero-Shot Learning uses previously acquired knowledge, semantic descriptions, attributes, or pretrained representations to handle target classes without labeled training examples from those classes. It does not necessarily mean that the model has never encountered related information.

3. What is the primary difference between Few-Shot Learning and Zero-Shot Learning?

  1. Few-Shot Learning uses a few labeled examples, while Zero-Shot Learning uses no task-specific labeled examples for the target classes
  2. Few-Shot Learning only works with text, while Zero-Shot Learning only works with images
  3. Few-Shot Learning cannot use pretrained models, while Zero-Shot Learning always requires them
  4. There is no difference between the two approaches

Answer: A) Few-Shot Learning uses a few labeled examples, while Zero-Shot Learning uses no task-specific labeled examples for the target classes

Explanation:

FSL adapts using a small number of labeled examples for a target task. ZSL relies on existing knowledge and relationships, such as textual descriptions or semantic embeddings, to recognize target categories without target-class labeled training examples.

4. What does the term "N-way K-shot" mean in Few-Shot Learning?

  1. N training epochs and K optimization steps
  2. N input features and K output classes
  3. N models trained with K datasets
  4. N classes with K labeled examples per class in the support set

Answer: D) N classes with K labeled examples per class in the support set

Explanation:

An N-way K-shot task contains N classes and K labeled examples per class in its support set. For example, a 5-way 2-shot task provides two labeled examples for each of five classes.

5. What is a support set in Few-Shot Learning?

  1. A collection of model checkpoints
  2. A small labeled dataset used to help the model adapt to a task
  3. A set containing only incorrect predictions
  4. A collection of unlabeled test results

Answer: B) A small labeled dataset used to help the model adapt to a task

Explanation:

The support set contains labeled examples that define the classes or task for a few-shot episode. The model uses these examples to establish class representations or adapt its predictions.

6. What is the purpose of a query set in episodic Few-Shot Learning?

  1. To store the model's source code
  2. To replace the training labels permanently
  3. To evaluate predictions on examples distinct from the support examples
  4. To define the number of model parameters

Answer: C) To evaluate predictions on examples distinct from the support examples

Explanation:

The query set contains examples used to assess how well the model generalizes from the support set. During episodic training, query losses can also provide learning signals for updating the model.

7. Which technique is strongly associated with learning how to adapt to new tasks from limited examples?

  1. Meta-learning
  2. Database normalization
  3. Static code analysis
  4. Lossless compression

Answer: A) Meta-learning

Explanation:

Meta-learning, often described as learning to learn, trains models across multiple tasks so they can adapt to new tasks with limited data. It is a common approach to Few-Shot Learning, although not all few-shot methods use meta-learning.

8. What is the main objective of Model-Agnostic Meta-Learning (MAML)?

  1. To eliminate the need for model parameters
  2. To classify only the categories seen during training
  3. To train a separate model for every possible future task
  4. To learn an initialization that can adapt to new tasks with a small number of gradient updates

Answer: D) To learn an initialization that can adapt to new tasks with a small number of gradient updates

Explanation:

MAML optimizes model parameters so that a small amount of task-specific training can produce good performance on a new task. It can be applied to few-shot classification and other learning problems.

9. Which method performs few-shot classification by comparing distances between learned class representations?

  1. Decision-tree pruning
  2. Prototypical Networks
  3. Apriori association mining
  4. Principal component reconstruction only

Answer: B) Prototypical Networks

Explanation:

Prototypical Networks calculate a prototype for each class, commonly by averaging the embeddings of its support examples. Query examples are classified according to their distances from the class prototypes.

10. How do Siamese Networks support Few-Shot Learning?

  1. By requiring a separate neural network for every input class
  2. By removing the need for feature representations
  3. By learning to compare examples using a similarity function or distance metric
  4. By converting every input into a database record

Answer: C) By learning to compare examples using a similarity function or distance metric

Explanation:

Siamese Networks use shared-weight branches to produce representations of input examples. A learned comparison function helps determine whether two examples are similar, which can be useful when only a few examples of a new class are available.

11. What is the role of transfer learning in Few-Shot Learning?

  1. It provides reusable knowledge learned from a source task or dataset
  2. It guarantees perfect predictions with one example
  3. It removes all pretrained parameters before training
  4. It prevents the model from recognizing new classes

Answer: A) It provides reusable knowledge learned from a source task or dataset

Explanation:

Transfer learning allows a model to reuse learned features or representations from a source task. This prior knowledge can make it easier to adapt to a new task when only a few labeled examples are available.

12. Which of the following is an example of a 3-way 1-shot learning task?

  1. Three classes with ten labeled examples per class
  2. One class with three labeled examples
  3. Three classes with three labeled examples per class
  4. Three classes with one labeled example per class

Answer: D) Three classes with one labeled example per class

Explanation:

A 3-way 1-shot task contains three classes and one labeled support example for each class. The model uses these three support examples to classify additional query examples.

13. What is a common strategy for training Few-Shot Learning models?

  1. Training only on the final test examples
  2. Using episodic training with support and query sets
  3. Removing labels from every training example
  4. Using a different label encoding for every prediction

Answer: B) Using episodic training with support and query sets

Explanation:

Episodic training simulates few-shot tasks during training. Each episode typically includes a support set and a query set, helping the model learn how to generalize from limited examples to additional examples from the same task.

14. Which problem can occur when a Few-Shot Learning model overfits its support set?

  1. The model automatically gains more training data
  2. The model becomes independent of input features
  3. The model performs well on support examples but poorly on unseen query examples
  4. The model can no longer calculate probabilities

Answer: C) The model performs well on support examples but poorly on unseen query examples

Explanation:

Overfitting occurs when a model adapts too closely to the small support set instead of learning patterns that generalize. Appropriate regularization, task diversity, and evaluation on separate query examples can help identify and reduce this problem.

15. Which metric is commonly used to evaluate a Few-Shot Classification model?

  1. Classification accuracy on query examples
  2. Number of lines in the training script
  3. Size of the model's variable names
  4. Number of comments in the source code

Answer: A) Classification accuracy on query examples

Explanation:

Query-set accuracy measures the proportion of query examples classified correctly. Depending on the task, precision, recall, F1-score, calibration, and confidence intervals across multiple episodes may also be important.

16. What does Zero-Shot Classification commonly use to connect an input with an unseen class?

  1. Only the number of training epochs
  2. Random class assignments without learned information
  3. Labels copied from unrelated test examples
  4. Semantic descriptions, attributes, or text representations of classes

Answer: D) Semantic descriptions, attributes, or text representations of classes

Explanation:

Zero-Shot Classification can use semantic information to relate an input to a class that lacks task-specific labeled training examples. For example, an image representation can be compared with text embeddings describing candidate categories.

17. What is a semantic embedding in Zero-Shot Learning?

  1. A random number assigned to each class
  2. A numerical representation of the meaning or attributes of an item or class
  3. A backup copy of a neural network
  4. A method for deleting training labels

Answer: B) A numerical representation of the meaning or attributes of an item or class

Explanation:

Semantic embeddings represent concepts, descriptions, or attributes as vectors. ZSL systems can use these representations to connect inputs with unseen classes based on learned semantic relationships.

18. How can a vision-language model perform Zero-Shot Image Classification?

  1. By requiring labeled training images for every candidate category
  2. By ignoring text descriptions entirely
  3. By comparing an image representation with text representations of candidate classes
  4. By assigning every image to the same class

Answer: C) By comparing an image representation with text representations of candidate classes

Explanation:

Vision-language models such as CLIP learn aligned image and text representations. At inference time, a system can compare an image embedding with embeddings of candidate class descriptions and select the best-matching description.

19. What is a common application of Zero-Shot Learning in Natural Language Processing?

  1. Classifying text into categories specified at inference time without task-specific labeled examples for those categories
  2. Increasing the physical memory of a computer
  3. Compiling source code into machine instructions
  4. Removing the semantic meaning of words

Answer: A) Classifying text into categories specified at inference time without task-specific labeled examples for those categories

Explanation:

Zero-shot text classification allows a model to evaluate text against candidate labels or descriptions that were not supplied as task-specific labeled training examples. For example, customer feedback can be classified into newly defined categories using a suitable pretrained model.

20. What does "unseen class" mean in a conventional Zero-Shot Learning setup?

  1. A class that does not have a name
  2. A class that cannot be represented numerically
  3. A class that is guaranteed to be impossible to predict
  4. A target class for which labeled training examples were not provided during the task's training phase

Answer: D) A target class for which labeled training examples were not provided during the task's training phase

Explanation:

An unseen class is excluded from the labeled training examples used for the target classification task. The model may still have learned related concepts or semantic knowledge that helps it recognize the class.

21. Which technique is frequently used to enable Zero-Shot Learning in large language models?

  1. Adding labeled examples for every target class to the prompt
  2. Providing instructions and a task description without task-specific examples
  3. Disabling the model's language understanding
  4. Training the model from scratch for every request

Answer: B) Providing instructions and a task description without task-specific examples

Explanation:

Zero-shot prompting supplies instructions and relevant context but no worked examples of the desired input-output behavior. A capable pretrained language model can use its existing learned patterns to attempt the task.

22. What is the primary difference between zero-shot prompting and few-shot prompting in a large language model?

  1. Zero-shot prompting requires model retraining, while few-shot prompting disables inference
  2. Zero-shot prompting works only with numerical data
  3. Zero-shot prompting provides instructions without demonstrations, while few-shot prompting includes a small number of demonstrations
  4. Few-shot prompting always requires thousands of examples

Answer: C) Zero-shot prompting provides instructions without demonstrations, while few-shot prompting includes a small number of demonstrations

Explanation:

Zero-shot prompts describe the task directly. Few-shot prompts include a few input-output examples that illustrate the desired pattern. Both approaches can use the same pretrained model without updating its parameters.

23. What is Generalized Zero-Shot Learning (GZSL)?

  1. An evaluation setting in which a model must recognize both seen and unseen classes
  2. A method that only predicts training classes
  3. A technique that requires every class to have identical images
  4. A learning method that cannot use semantic attributes

Answer: A) An evaluation setting in which a model must recognize both seen and unseen classes

Explanation:

Generalized Zero-Shot Learning evaluates a model on both previously seen and unseen classes. This is more challenging than conventional ZSL because models may favor familiar classes when making predictions.

24. Which issue is commonly encountered in Generalized Zero-Shot Learning?

  1. All classes automatically have the same number of samples
  2. Unseen classes always receive higher scores than seen classes
  3. Semantic descriptions cannot be stored as vectors
  4. The model may be biased toward classes seen during training

Answer: D) The model may be biased toward classes seen during training

Explanation:

Models often favor seen classes because their representations are directly supported by labeled training data. GZSL methods may address this bias through calibration, training objectives, or other techniques that improve the balance between seen and unseen class predictions.

25. What is the role of class attributes in attribute-based Zero-Shot Learning?

  1. They determine the number of neural network layers
  2. They describe properties that help relate unseen classes to learned concepts
  3. They replace every input feature with a random value
  4. They prevent the model from predicting unseen categories

Answer: B) They describe properties that help relate unseen classes to learned concepts

Explanation:

Class attributes can describe characteristics such as color, shape, texture, or behavior. A model can learn relationships between input features and attributes, then use attribute descriptions to recognize classes that were not represented by labeled training examples.

26. Which of the following best describes the relationship between FSL and transfer learning?

  1. FSL and transfer learning are unrelated in every application
  2. Transfer learning always eliminates the need for target-task data
  3. FSL can use representations learned through transfer learning to adapt with fewer labeled examples
  4. FSL requires discarding all pretrained knowledge

Answer: C) FSL can use representations learned through transfer learning to adapt with fewer labeled examples

Explanation:

Transfer learning provides useful representations or initialization from previous training. Few-Shot Learning can build on this knowledge to perform a new task with limited labeled examples, reducing dependence on large task-specific datasets.

27. Why are pretrained models useful for Zero-Shot Learning?

  1. They can contain reusable knowledge and semantic relationships learned from broad training data
  2. They guarantee that every unseen class will be classified correctly
  3. They eliminate the need to define a task
  4. They automatically create ground-truth labels for every possible input

Answer: A) They can contain reusable knowledge and semantic relationships learned from broad training data

Explanation:

Pretrained models can transfer knowledge from their original training tasks to new tasks. This prior knowledge can support zero-shot predictions, although performance depends on the model, input distribution, task design, and quality of the descriptions.

28. What is the purpose of a similarity score in embedding-based Zero-Shot Learning?

  1. To measure the physical size of the training dataset
  2. To count the number of model layers
  3. To calculate how many labels were manually created
  4. To estimate how closely an input representation matches a candidate class representation

Answer: D) To estimate how closely an input representation matches a candidate class representation

Explanation:

A similarity score, such as cosine similarity, compares vector representations of an input and candidate classes. The model can use these scores to rank candidates, although high similarity does not always guarantee a correct prediction.

29. What is cosine similarity commonly used for in embedding-based FSL or ZSL systems?

  1. Measuring the number of training batches
  2. Comparing the angular similarity between vector representations
  3. Calculating the storage capacity of a GPU
  4. Determining the number of labels in a dataset without inspecting it

Answer: B) Comparing the angular similarity between vector representations

Explanation:

Cosine similarity measures the cosine of the angle between two nonzero vectors. It is frequently used to compare semantic embeddings because it emphasizes their directional alignment rather than their magnitudes alone.

30. Which of the following is a limitation of Few-Shot Learning?

  1. It cannot be applied to image classification
  2. It always requires every class to have millions of examples
  3. Its performance may be sensitive to the quality and representativeness of the few labeled examples
  4. It cannot use pretrained feature extractors

Answer: C) Its performance may be sensitive to the quality and representativeness of the few labeled examples

Explanation:

Few-shot tasks provide limited labeled data, so mislabeled, unrepresentative, or ambiguous examples can strongly affect predictions. The model's prior knowledge and adaptation strategy also influence performance.

31. Which of the following is a limitation of Zero-Shot Learning?

  1. It can struggle when the target classes are poorly described or differ substantially from the model's learned knowledge
  2. It always requires a separate labeled example for every unseen class
  3. It cannot use textual class descriptions
  4. It only supports binary classification

Answer: A) It can struggle when the target classes are poorly described or differ substantially from the model's learned knowledge

Explanation:

ZSL depends on the model's prior knowledge and the information used to represent target classes. Weak descriptions, domain shifts, ambiguous labels, and unfamiliar concepts can reduce prediction quality.

32. What is a major advantage of Few-Shot Learning in medical image analysis?

  1. It removes the need for medical expertise in every workflow
  2. It guarantees perfect diagnosis from a single image
  3. It prevents models from detecting rare conditions
  4. It can help adapt a model to a rare condition when only a small number of labeled examples are available

Answer: D) It can help adapt a model to a rare condition when only a small number of labeled examples are available

Explanation:

Rare medical conditions may have limited labeled datasets. FSL can help transfer knowledge from related cases to a new classification task, but clinical validation, representative data, and expert oversight remain essential.

33. How can Zero-Shot Learning support image recognition for new product categories?

  1. By retraining a complete model for every new product image
  2. By comparing image features with textual descriptions of the new product categories
  3. By assigning all new products to a single existing category
  4. By removing product descriptions from the classification process

Answer: B) By comparing image features with textual descriptions of the new product categories

Explanation:

A vision-language model can compare product images with text descriptions of candidate categories. This enables recognition of new categories without requiring a separate labeled image training set for each category.

34. Which statement about Few-Shot Learning and model fine-tuning is correct?

  1. Few-Shot Learning always prohibits parameter updates
  2. Fine-tuning always requires millions of labeled examples
  3. Few-Shot Learning may use limited-example fine-tuning or non-parametric methods that do not update model parameters
  4. Fine-tuning and inference are identical operations

Answer: C) Few-Shot Learning may use limited-example fine-tuning or non-parametric methods that do not update model parameters

Explanation:

Some FSL methods update model parameters using a small support set, while others classify examples using distances to support examples or prototypes without further gradient-based updates. The appropriate strategy depends on the model and task.

35. What does a "class prototype" represent in Prototypical Networks?

  1. A representative embedding, often calculated as the mean of a class's support-example embeddings
  2. The largest possible neural network in a model family
  3. A list of all test-set labels
  4. A manually selected optimization algorithm

Answer: A) A representative embedding, often calculated as the mean of a class's support-example embeddings

Explanation:

A class prototype summarizes the embeddings of labeled support examples belonging to that class. Query examples can then be classified according to their distances from the available prototypes.

36. Why is the separation of training, validation, and test classes important when evaluating Few-Shot Learning?

  1. It makes all classes identical
  2. It removes the need for an evaluation metric
  3. It ensures that the model memorizes the test examples
  4. It helps measure generalization to classes or tasks not used to train the model's underlying parameters

Answer: D) It helps measure generalization to classes or tasks not used to train the model's underlying parameters

Explanation:

In standard class-disjoint few-shot evaluation, training, validation, and test classes are separated to assess generalization to new categories. Data leakage between these splits can make reported performance misleading.

37. Which statement best describes in-context learning in large language models?

  1. It always requires retraining the entire model
  2. The model uses instructions or demonstrations in the prompt to perform a task without necessarily updating its parameters
  3. It works only when the prompt contains no text
  4. It permanently stores every prompt as a new model parameter

Answer: B) The model uses instructions or demonstrations in the prompt to perform a task without necessarily updating its parameters

Explanation:

In-context learning uses information supplied within the current context window. A zero-shot prompt may contain instructions only, while a few-shot prompt includes demonstrations that guide the model's response.

38. What is prompt sensitivity in Zero-Shot Learning with large language models?

  1. The model's ability to measure the physical size of a prompt
  2. A guarantee that all prompt formats produce identical results
  3. Changes in model output caused by differences in prompt wording, structure, or context
  4. The process of removing the prompt before inference

Answer: C) Changes in model output caused by differences in prompt wording, structure, or context

Explanation:

Zero-shot performance can vary with the instructions, label descriptions, and context provided to a model. Testing multiple prompt formulations and evaluating on representative examples can help identify this sensitivity.

39. What is the purpose of class calibration in some Zero-Shot Classification systems?

  1. To adjust systematic prediction biases or score differences among candidate classes
  2. To guarantee that every prediction is correct
  3. To increase the number of labeled training examples automatically
  4. To convert every unseen class into a seen class

Answer: A) To adjust systematic prediction biases or score differences among candidate classes

Explanation:

Calibration methods can help correct systematic preferences for certain classes or improve the interpretation of prediction scores. In generalized ZSL, calibration is often used to address the tendency to favor seen classes over unseen ones.

40. Which evaluation measure is particularly useful when classifying an imbalanced dataset using FSL or ZSL?

  1. Number of training files
  2. Length of the input feature names
  3. Total number of model parameters alone
  4. Macro-averaged F1-score

Answer: D) Macro-averaged F1-score

Explanation:

Macro-averaged F1 calculates the F1-score for each class and averages the scores with equal class weight. It can reveal poor performance on minority classes that overall accuracy may hide.

41. Which dataset design is appropriate for evaluating a Zero-Shot Learning image classifier?

  1. Include labeled training images from every target class without restrictions
  2. Train on designated seen classes and evaluate on separate target classes represented through suitable semantic descriptions
  3. Evaluate only on the exact images used to train the model
  4. Remove all class descriptions before training and testing every model in the same way

Answer: B) Train on designated seen classes and evaluate on separate target classes represented through suitable semantic descriptions

Explanation:

A conventional ZSL benchmark separates seen training classes from unseen evaluation classes and supplies semantic information for the target classes. This tests whether the model can transfer learned relationships to classes without labeled training examples.

42. How can Few-Shot Learning help with personalized recommendation systems?

  1. By guaranteeing that new users never receive irrelevant recommendations
  2. By requiring every user to interact with thousands of items first
  3. By adapting recommendations from a small amount of interaction data for a new user or context
  4. By preventing the system from using previous user interactions

Answer: C) By adapting recommendations from a small amount of interaction data for a new user or context

Explanation:

Few-Shot Learning can help recommendation systems adapt when a new user has provided only a few interactions. Its success depends on the quality of the prior representations, the relevance of the available interactions, and the evaluation setting.

43. Which statement about Zero-Shot Learning and unseen categories is correct?

  1. The model can predict only categories that have labeled examples in the task-specific training set
  2. The model must receive at least one labeled example for every target category
  3. Semantic information guarantees perfect recognition of unseen categories
  4. The model can attempt to recognize unseen categories when useful semantic relationships or pretrained knowledge are available

Answer: D) The model can attempt to recognize unseen categories when useful semantic relationships or pretrained knowledge are available

Explanation:

ZSL aims to generalize to categories without task-specific labeled training examples. Semantic descriptions and pretrained representations can provide a bridge between previously learned concepts and new categories, but they do not guarantee correct predictions.

44. What is a key difference between Few-Shot Learning and conventional supervised learning with a large labeled dataset?

  1. Few-Shot Learning is designed to work with very limited labeled examples for the target task
  2. Few-Shot Learning never uses labels
  3. Conventional supervised learning cannot classify images
  4. Few-Shot Learning does not require any evaluation

Answer: A) Few-Shot Learning is designed to work with very limited labeled examples for the target task

Explanation:

Conventional supervised learning commonly relies on a substantial labeled dataset for each target task. Few-Shot Learning focuses on learning or adapting from a small number of labeled examples, often by exploiting prior knowledge.

45. Which approach can improve Few-Shot Learning when only a few labeled examples are available?

  1. Using only random labels
  2. Applying data augmentation where appropriate and using pretrained representations
  3. Discarding relevant training knowledge
  4. Evaluating only on the support set

Answer: B) Applying data augmentation where appropriate and using pretrained representations

Explanation:

Suitable data augmentation can increase the diversity of training inputs, while pretrained representations provide reusable features. These methods can improve generalization, although inappropriate augmentation or domain mismatch can reduce performance.

46. What is the role of text descriptions in a vision-language Zero-Shot Learning system?

  1. They replace the image input in every task
  2. They guarantee that visually similar classes can always be distinguished
  3. They provide semantic representations that can be compared with image representations
  4. They eliminate the need to evaluate the model

Answer: C) They provide semantic representations that can be compared with image representations

Explanation:

Text descriptions encode candidate class meanings in a form that can be aligned with image features. The model can compare the representations to rank possible labels, though vague or overlapping descriptions can cause confusion.

47. Which statement best describes the relationship between Few-Shot Learning and Zero-Shot Learning?

  1. Both aim to generalize beyond large amounts of task-specific labeled data, but they differ in whether a few labeled target-task examples are provided
  2. Both require exactly one million labeled examples per class
  3. Both are limited to language models and cannot support computer vision
  4. Both guarantee the same performance on every dataset

Answer: A) Both aim to generalize beyond large amounts of task-specific labeled data, but they differ in whether a few labeled target-task examples are provided

Explanation:

FSL uses a small number of labeled examples for a target task, whereas ZSL attempts to handle target classes without task-specific labeled examples for those classes. Both can exploit pretrained knowledge, semantic relationships, and transferable representations.

48. A developer needs to classify images of a newly discovered bird species but has only three labeled images. Which approach is most directly suited to this situation?

  1. Training a large image classifier from scratch using only the three images
  2. Rejecting the new species because its dataset is small
  3. Using only a database query without image representations
  4. Applying Few-Shot Learning with a pretrained image representation and the available labeled examples

Answer: D) Applying Few-Shot Learning with a pretrained image representation and the available labeled examples

Explanation:

Few-Shot Learning is suited to adapting a model with a small number of labeled examples. A pretrained image encoder can provide useful features, while a suitable adaptation or prototype-based method can support classification. The system should still be tested on independent examples.

49. A company wants to sort customer support messages into a newly introduced category, but it has no labeled training examples for that category. The category can be clearly described in natural language. Which approach is most appropriate to try first?

  1. Train a classifier from scratch without any training data
  2. Use Zero-Shot Text Classification with a suitable pretrained language model and a clear category description
  3. Assign every message to the new category
  4. Remove the category description before inference

Answer: B) Use Zero-Shot Text Classification with a suitable pretrained language model and a clear category description

Explanation:

Zero-Shot Text Classification can evaluate messages against a category described at inference time without labeled examples for that target category. The company should validate predictions against manually reviewed messages before relying on them in a production workflow.

50. An AI research team must classify images into both familiar animal classes and new animal classes that have no labeled training images. It has a pretrained vision-language model, textual descriptions of all candidate classes, and a small labeled support set for some new classes. Which strategy best addresses the requirements?

  1. Train a separate image classifier from scratch for every class and discard the pretrained model
  2. Use only the support set and ignore all class descriptions and pretrained knowledge
  3. Use the vision-language model for zero-shot predictions, adapt or calibrate predictions with the available support examples where appropriate, and evaluate performance on both seen and unseen classes
  4. Assume that textual descriptions guarantee perfect predictions and skip testing

Answer: C) Use the vision-language model for zero-shot predictions, adapt or calibrate predictions with the available support examples where appropriate, and evaluate performance on both seen and unseen classes

Explanation:

This scenario combines zero-shot recognition with limited-example adaptation. Textual descriptions and pretrained image-text representations support predictions for classes without labeled training images, while the available support examples can help adapt or calibrate predictions for selected new classes. Evaluation should include both familiar and new classes to measure generalization and detect bias toward seen categories.

Comments and Discussions!

Load comments ↻



Top MCQs
Top Programs/Examples
Top Tutorials
About
Student's Section
Subscribe

Copyright © 2026 www.includehelp.com. All rights reserved.