Home »
Trending Technologies MCQs
Meta-Learning MCQs (Multiple-Choice Questions)
Practice Meta-Learning MCQs to test your understanding of machine learning techniques that enable models to learn new tasks quickly by leveraging experience from previous tasks. These multiple-choice questions cover learning to learn, meta-training, meta-testing, optimization-based methods, metric-based methods, and memory-based approaches. 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 Meta-Learning.
Meta-Learning MCQs
These Meta-Learning multiple-choice questions cover important concepts such as task distributions, episodic training, few-shot learning, Model-Agnostic Meta-Learning (MAML), Reptile, Prototypical Networks, optimization-based learning, metric-based learning, meta-gradients, and adaptation strategies. The questions explore how models use experience across multiple tasks to adapt to unfamiliar problems with limited data through conceptual, technical, and scenario-based questions.
These Meta-Learning MCQs help learners understand how models acquire reusable knowledge, optimize their ability to adapt, and generalize to new tasks with fewer examples. Each question includes an answer and explanation.
List of Meta-Learning MCQs
Explore the following 50 MCQs covering Meta-Learning fundamentals, learning paradigms, popular algorithms, applications, evaluation methods, challenges, and practical implementation scenarios.
1. What is Meta-Learning in machine learning?
- A technique that trains a model on only one fixed task
- A method used exclusively for storing training datasets
- An approach in which a model learns how to learn new tasks using experience from previous tasks
- A process that removes all learned parameters before inference
Answer: C) An approach in which a model learns how to learn new tasks using experience from previous tasks
Explanation:
Meta-Learning is often described as learning to learn. Instead of optimizing only for one task, a meta-learning system uses experience across multiple tasks to improve its ability to adapt to new tasks with limited data or computation.
2. What is the primary objective of Meta-Learning?
- To enable rapid adaptation to new tasks using knowledge acquired from previous tasks
- To memorize every training example without generalization
- To eliminate the need for optimization in all learning systems
- To ensure that every task uses identical input data
Answer: A) To enable rapid adaptation to new tasks using knowledge acquired from previous tasks
Explanation:
Meta-Learning aims to improve how a model learns across tasks. Depending on the method, it may learn a suitable initialization, a distance metric, an adaptation strategy, or a memory mechanism that supports learning from limited examples.
3. Which phrase is commonly associated with Meta-Learning?
- Learning without any input
- Learning only through memorization
- Training without evaluating performance
- Learning to learn
Answer: D) Learning to learn
Explanation:
The phrase "learning to learn" describes the central idea of Meta-Learning. A model learns from a collection of tasks so that it can become better at adapting to other tasks rather than treating each new problem as entirely independent.
4. What is a meta-training task?
- A task that can only be completed after deployment
- A task used during meta-training to help the model learn a transferable learning strategy
- A task that contains no data or objective
- A task used only to measure hardware performance
Answer: B) A task used during meta-training to help the model learn a transferable learning strategy
Explanation:
Meta-training exposes the model to multiple tasks. The training process uses feedback from these tasks to learn knowledge or parameters that can support adaptation to new tasks at meta-test time.
5. What is the role of meta-testing in Meta-Learning?
- To retrain the entire system on every possible task
- To remove all previously learned information
- To evaluate how well the learned strategy adapts to new tasks
- To generate random labels for the training set
Answer: C) To evaluate how well the learned strategy adapts to new tasks
Explanation:
Meta-testing measures generalization to tasks not used during meta-training. A model may receive a small support set for each test task, adapt according to its learned strategy, and then be evaluated on separate query examples.
6. Which learning problem is closely associated with Meta-Learning?
- Few-Shot Learning
- File compression
- Database indexing
- Static website rendering
Answer: A) Few-Shot Learning
Explanation:
Few-Shot Learning requires a model to learn a new task from a small number of labeled examples. Meta-Learning can help by training the model across related tasks so it can adapt more effectively when only a few examples are available.
7. What is episodic training in Meta-Learning?
- Training a model continuously on one unchanging example
- Removing all labels before model training
- Evaluating a model only after deployment
- Training with task-like episodes that simulate the adaptation and evaluation process
Answer: D) Training with task-like episodes that simulate the adaptation and evaluation process
Explanation:
Episodic training organizes data into simulated tasks, often with support and query sets. This allows the model to practice adapting to a task and being evaluated on additional examples, which resembles its intended use on new tasks.
8. What is Model-Agnostic Meta-Learning (MAML)?
- A method that works only with decision trees
- A meta-learning algorithm that learns an initialization suitable for rapid adaptation through a small number of gradient updates
- A method that prevents model parameters from changing during every training process
- A database technique for managing neural network checkpoints
Answer: B) A meta-learning algorithm that learns an initialization suitable for rapid adaptation through a small number of gradient updates
Explanation:
MAML optimizes initial model parameters so that a small amount of task-specific training can produce good performance on a new task. Its general formulation can be applied to different differentiable model architectures and learning objectives.
9. Why is MAML described as model-agnostic?
- It does not require any model parameters
- It can only be used with linear regression
- Its general approach can be applied to different differentiable models trained using gradient-based optimization
- It guarantees the same accuracy for every architecture
Answer: C) Its general approach can be applied to different differentiable models trained using gradient-based optimization
Explanation:
MAML is not tied to one specific neural network architecture. Its core method optimizes an initialization through task-specific gradient updates, although practical implementations require differentiable models and appropriate optimization procedures.
10. What is the purpose of the inner loop in MAML?
- To adapt the model parameters to an individual task using that task's support data
- To permanently delete the model's initialization
- To evaluate unrelated tasks without using their data
- To calculate the storage capacity of the training system
Answer: A) To adapt the model parameters to an individual task using that task's support data
Explanation:
The inner loop performs task-specific learning using the support examples of an episode. MAML then evaluates the adapted parameters on query examples to obtain information used to improve the initial parameters.
11. What is the purpose of the outer loop in MAML?
- To replace all training tasks with a single fixed task
- To freeze all model parameters permanently
- To select the model's file format
- To update the initial parameters based on performance after task-specific adaptation
Answer: D) To update the initial parameters based on performance after task-specific adaptation
Explanation:
The outer loop uses query-set performance from multiple tasks to update the meta-initialization. It aims to find parameters from which the model can adapt effectively to new tasks.
12. What is Reptile in Meta-Learning?
- A reinforcement learning environment for robotic animals
- A first-order meta-learning algorithm that moves the initialization toward parameters obtained after task-specific training
- A method for converting neural networks into database tables
- A classification metric used only for imbalanced datasets
Answer: B) A first-order meta-learning algorithm that moves the initialization toward parameters obtained after task-specific training
Explanation:
Reptile samples tasks, trains model parameters on each task, and moves the shared initialization toward the resulting task-trained parameters. It provides a simpler alternative to full second-order MAML optimization.
13. What is a key difference between MAML and Reptile?
- Reptile cannot use neural networks
- MAML never uses task-specific training
- MAML commonly optimizes query loss through the adaptation process, while Reptile uses a simpler update toward task-trained parameters
- Both algorithms require identical mathematical update rules
Answer: C) MAML commonly optimizes query loss through the adaptation process, while Reptile uses a simpler update toward task-trained parameters
Explanation:
Standard MAML differentiates through inner-loop updates to optimize post-adaptation performance. Reptile uses a first-order update based on the difference between the initialization and parameters learned on sampled tasks, avoiding the full second-order calculation used by standard MAML.
14. What is a metric-based Meta-Learning method?
- A method that learns a similarity measure or embedding space for comparing examples
- A method that measures only the physical size of a dataset
- A technique that stores every prediction in a relational database
- A method that removes all input features before classification
Answer: A) A method that learns a similarity measure or embedding space for comparing examples
Explanation:
Metric-based methods learn representations in which examples from the same class or related tasks can be compared meaningfully. During adaptation, a new example may be classified using distances or similarities to a small support set.
15. Which algorithm is a well-known metric-based few-shot learning method?
- Quick sort
- Prototypical Networks
- Binary search
- Database normalization
Answer: B) Prototypical Networks
Explanation:
Prototypical Networks learn an embedding space and represent each class with a prototype, commonly the mean of its support-example embeddings. Query examples are classified according to their distances from these prototypes.
16. How do Prototypical Networks usually calculate a class prototype?
- By selecting a random output label
- By counting the number of neural network layers
- By choosing the largest training file
- By averaging the embeddings of support examples belonging to the class
Answer: D) By averaging the embeddings of support examples belonging to the class
Explanation:
A prototype is a representative vector for a class. In standard Prototypical Networks, the embeddings of the labeled support examples for that class are averaged, and query examples are classified based on distances to the prototypes.
17. What is a Siamese Network commonly used for in Meta-Learning-related applications?
- Learning to compare pairs of examples using shared-weight neural network branches
- Increasing storage capacity without processing data
- Removing similarity information from embeddings
- Generating labels without defining any objective
Answer: A) Learning to compare pairs of examples using shared-weight neural network branches
Explanation:
Siamese Networks process two inputs through branches that share parameters and learn a similarity or distance function. This can support one-shot and few-shot recognition when examples of a new category are scarce.
18. What is optimization-based Meta-Learning?
- A technique that learns only class names
- A process that avoids using any training objective
- A family of methods that learns parameters or update strategies to support rapid task adaptation
- A method used exclusively for sorting numeric arrays
Answer: C) A family of methods that learns parameters or update strategies to support rapid task adaptation
Explanation:
Optimization-based Meta-Learning includes methods such as MAML that optimize an initialization or learning process across multiple tasks. The goal is to make task-specific learning more effective when a new task arrives.
19. What is memory-based Meta-Learning?
- A method that stores only model filenames
- An approach that uses an external or internal memory mechanism to retain and apply information across examples or tasks
- A method that prevents a model from using previous examples
- A database backup technique unrelated to learning
Answer: B) An approach that uses an external or internal memory mechanism to retain and apply information across examples or tasks
Explanation:
Memory-based approaches allow a model to store or retrieve information useful for making predictions or adapting to new tasks. The memory may be implemented through a differentiable external memory, learned state, or other architecture-specific mechanisms.
20. What is the role of a task distribution in Meta-Learning?
- It determines the color of training visualizations
- It guarantees that all tasks have identical labels
- It removes the need for separate training and testing tasks
- It describes how tasks are sampled for meta-training and the kinds of tasks to which the model is expected to generalize
Answer: D) It describes how tasks are sampled for meta-training and the kinds of tasks to which the model is expected to generalize
Explanation:
Meta-Learning typically assumes access to a collection or distribution of tasks. The model learns across sampled tasks, and its performance depends partly on whether the evaluation tasks are sufficiently related to the training task distribution.
21. What is a support set in episodic Meta-Learning?
- A small collection of labeled examples used to adapt to an episode's task
- A set of software installation files
- A collection of unrelated model architectures
- A list containing only incorrect predictions
Answer: A) A small collection of labeled examples used to adapt to an episode's task
Explanation:
The support set provides the examples available for task-specific adaptation. In few-shot classification, it commonly contains a small number of labeled examples from each class in the current episode.
22. What is a query set in episodic Meta-Learning?
- A collection of model configuration files
- A set used to determine the computer's processor speed
- A set of examples used to evaluate predictions after adaptation and often provide the meta-training loss
- A collection of support examples duplicated without evaluation
Answer: C) A set of examples used to evaluate predictions after adaptation and often provide the meta-training loss
Explanation:
Query examples are distinct from the support examples used for adaptation. Their predictions measure post-adaptation generalization and, in many meta-training methods, contribute to the objective used to update meta-learned parameters.
23. What does N-way K-shot mean in few-shot Meta-Learning?
- N training epochs and K optimizers
- N input features and K hidden layers
- N models and K deployment servers
- N classes with K labeled support examples per class
Answer: D) N classes with K labeled support examples per class
Explanation:
An N-way K-shot task contains N target classes and K labeled support examples per class. For example, a 5-way 1-shot episode contains five classes with one support example for each class.
24. What is the purpose of meta-gradients in gradient-based Meta-Learning?
- To measure the physical size of the training dataset
- To provide information for updating meta-level parameters based on task performance after adaptation
- To convert training labels into images
- To remove all optimization steps from the model
Answer: B) To provide information for updating meta-level parameters based on task performance after adaptation
Explanation:
Meta-gradients guide updates to the parameters learned across tasks. In MAML, they are commonly calculated by differentiating the query loss through the inner-loop adaptation steps, although first-order approximations can reduce computational cost.
25. Which is a major advantage of Meta-Learning?
- It can help models adapt to new tasks using fewer task-specific examples
- It guarantees perfect performance on every unseen task
- It eliminates all computational requirements
- It prevents models from learning transferable representations
Answer: A) It can help models adapt to new tasks using fewer task-specific examples
Explanation:
Meta-Learning can improve sample efficiency by learning from multiple related tasks. This is especially useful when labeled examples are expensive or difficult to collect, though successful adaptation depends on the quality and diversity of meta-training tasks.
26. What is a common challenge in Meta-Learning?
- It can only process text data
- It does not use any training data
- It may generalize poorly when new tasks differ substantially from the meta-training tasks
- It cannot use optimization algorithms
Answer: C) It may generalize poorly when new tasks differ substantially from the meta-training tasks
Explanation:
A meta-learned strategy is influenced by the tasks encountered during meta-training. If deployment tasks have very different distributions, labels, or objectives, the learned adaptation strategy may no longer be suitable.
27. How does Meta-Learning differ from conventional machine learning?
- Conventional machine learning never uses data
- Meta-Learning focuses on improving learning or adaptation across tasks, while conventional training often optimizes directly for a particular task
- Meta-Learning can only be performed without models
- Conventional machine learning always requires more tasks than Meta-Learning
Answer: B) Meta-Learning focuses on improving learning or adaptation across tasks, while conventional training often optimizes directly for a particular task
Explanation:
Conventional supervised learning usually minimizes a task-specific objective using examples from that task. Meta-Learning optimizes across multiple tasks so that a model can learn new tasks more effectively or adapt with limited data.
28. Which type of Meta-Learning learns a parameter initialization that can be adapted quickly?
- Rule-based sorting
- Database indexing
- Static code analysis
- Optimization-based Meta-Learning
Answer: D) Optimization-based Meta-Learning
Explanation:
Optimization-based methods such as MAML learn an initialization that can be adjusted using a small number of gradient steps on a new task. The initialization is optimized based on how well the adapted model performs across training tasks.
29. What is the purpose of task diversity during meta-training?
- To expose the model to varied tasks so it can learn a more general adaptation strategy
- To make every training episode identical
- To prevent the model from encountering new examples
- To eliminate the need for task-specific evaluation
Answer: A) To expose the model to varied tasks so it can learn a more general adaptation strategy
Explanation:
Diverse meta-training tasks can help a model learn patterns that generalize beyond one specific task. However, the tasks must still be relevant to the intended application; excessive mismatch can make useful adaptation more difficult.
30. Which evaluation metric is commonly used for few-shot classification in Meta-Learning?
- Number of model filenames
- Accuracy on query examples across multiple test episodes
- Number of comments in the training code
- Size of the computer monitor
Answer: B) Accuracy on query examples across multiple test episodes
Explanation:
Query accuracy measures the proportion of examples classified correctly after adaptation. Reporting the mean across multiple episodes, along with a measure of uncertainty, helps characterize performance across different sampled tasks.
31. What is the purpose of separating meta-training and meta-testing tasks?
- To make every task use the same examples
- To prevent the model from learning any shared information
- To evaluate generalization to tasks that were not used to learn the meta-level strategy
- To remove the need for a validation process
Answer: C) To evaluate generalization to tasks that were not used to learn the meta-level strategy
Explanation:
Separate meta-training and meta-testing tasks help determine whether the model has learned a transferable adaptation strategy rather than simply memorizing training tasks. Careful task and data separation is important for a reliable evaluation.
32. How can Meta-Learning be applied to robotics?
- By replacing robot sensors with random numbers
- By removing all feedback from control systems
- By preventing robots from adapting to new environments
- By helping a robot adapt its control policy to new conditions or tasks using experience from previous tasks
Answer: D) By helping a robot adapt its control policy to new conditions or tasks using experience from previous tasks
Explanation:
A robot may encounter changes in terrain, object properties, or operating conditions. Meta-Learning can help it adapt using limited new experience, potentially reducing the amount of retraining needed for each environment.
33. Why is Meta-Learning useful for personalized recommendation systems?
- It can help adapt recommendations to new users using limited interaction data
- It guarantees that every user prefers the same items
- It eliminates the need to evaluate recommendations
- It prevents the system from learning from previous users or tasks
Answer: A) It can help adapt recommendations to new users using limited interaction data
Explanation:
Meta-Learning can train a recommendation system across multiple user-related tasks so that it adapts to a new user from a small number of interactions. Its effectiveness depends on task design, user diversity, and the quality of the interaction data.
34. What is a limitation of second-order MAML?
- It cannot be used with gradient-based learning
- It can require substantial computation and memory because it differentiates through inner-loop optimization
- It never uses a support set
- It cannot be evaluated on unseen tasks
Answer: B) It can require substantial computation and memory because it differentiates through inner-loop optimization
Explanation:
Standard second-order MAML computes meta-gradients that account for the dependence of adapted parameters on the initialization. Differentiating through inner-loop updates can be expensive, so first-order approximations and other algorithms are sometimes used.
35. What is First-Order MAML (FOMAML)?
- A method that does not use any gradient information
- A technique that trains only a nearest-neighbor classifier
- An approximation to MAML that ignores certain second-order derivative terms
- A method that requires no task-specific adaptation
Answer: C) An approximation to MAML that ignores certain second-order derivative terms
Explanation:
First-Order MAML simplifies the meta-gradient computation by ignoring second-order derivative terms. This can reduce computational cost while retaining the general strategy of learning an initialization that adapts to new tasks.
36. Which statement about Meta-Learning and Transfer Learning is correct?
- Transfer Learning never uses pretrained models
- Meta-Learning and Transfer Learning cannot be combined
- Transfer Learning applies only to image data
- Meta-Learning can learn an adaptation strategy across tasks, while Transfer Learning reuses knowledge from a source task or dataset for a target task
Answer: D) Meta-Learning can learn an adaptation strategy across tasks, while Transfer Learning reuses knowledge from a source task or dataset for a target task
Explanation:
Transfer Learning often initializes a model with pretrained knowledge and adapts it to a target task. Meta-Learning focuses on learning how to adapt across a distribution of tasks. A meta-learned initialization can itself be transferred to new tasks.
37. What role can reinforcement learning play in Meta-Learning?
- It can provide tasks and reward signals for learning how to adapt policies across different environments
- It guarantees that a policy never requires further training
- It eliminates the need for an environment or reward definition
- It converts all reinforcement learning problems into image classification tasks
Answer: A) It can provide tasks and reward signals for learning how to adapt policies across different environments
Explanation:
Meta-Reinforcement Learning studies how an agent can learn to adapt its behavior across tasks or environments. The meta-training process can use experience and reward signals to develop policies or learning strategies that perform better on new tasks.
38. What is a common risk when meta-training and meta-testing use overlapping examples or tasks?
- The model automatically becomes more robust to every domain
- Data leakage can lead to overly optimistic estimates of generalization
- The model can no longer calculate a loss
- The optimizer stops updating parameters
Answer: B) Data leakage can lead to overly optimistic estimates of generalization
Explanation:
If test tasks or examples influence meta-training or model selection improperly, evaluation results may not reflect performance on genuinely unseen tasks. Separate task splits and careful dataset construction help reduce this risk.
39. Which method is commonly associated with memory-based meta-learning architectures?
- Merge sort
- Binary search tree balancing
- Memory-Augmented Neural Networks
- Database table normalization
Answer: C) Memory-Augmented Neural Networks
Explanation:
Memory-Augmented Neural Networks use a neural network together with a memory mechanism that can store and retrieve information. Such architectures can support rapid learning from new examples by retaining information useful for subsequent predictions.
40. What is the purpose of regularization in Meta-Learning?
- To force every model parameter to become zero
- To guarantee identical predictions for all tasks
- To eliminate the need for training data
- To help control overfitting and encourage generalization across tasks
Answer: D) To help control overfitting and encourage generalization across tasks
Explanation:
Regularization can discourage a model from becoming too specialized to the meta-training tasks. Depending on the method, it may constrain parameters, encourage simpler solutions, or improve robustness when adapting to new tasks.
41. How can Meta-Learning help in medical image classification?
- By helping a model adapt to a new diagnostic classification task with a small number of labeled examples
- By eliminating the need for medical validation
- By guaranteeing that every diagnosis is correct
- By replacing all medical images with random data
Answer: A) By helping a model adapt to a new diagnostic classification task with a small number of labeled examples
Explanation:
Medical datasets for rare conditions may contain relatively few labeled images. Meta-Learning can help models adapt to such tasks by learning from related training tasks, but clinical validation and expert oversight remain essential.
42. What is the purpose of hyperparameter selection in Meta-Learning?
- To ensure that the model uses the same learning rate for every possible task
- To select settings such as inner-loop learning rate, number of adaptation steps, and meta-learning rate
- To remove the need for validation tasks
- To determine class labels without examining the data
Answer: B) To select settings such as inner-loop learning rate, number of adaptation steps, and meta-learning rate
Explanation:
Meta-Learning has configuration choices that affect both meta-training and task adaptation. Validation tasks can help determine suitable learning rates, adaptation steps, architecture settings, and other hyperparameters without using the final test tasks for repeated selection.
43. What is the relationship between Meta-Learning and generalization?
- Meta-Learning focuses only on memorizing training labels
- Generalization is irrelevant when a model learns across tasks
- Meta-Learning aims to improve a model's ability to adapt and perform well on new tasks beyond those used for meta-training
- Meta-Learning guarantees identical performance on every unseen task
Answer: C) Meta-Learning aims to improve a model's ability to adapt and perform well on new tasks beyond those used for meta-training
Explanation:
A key goal of Meta-Learning is task-level generalization. A successful meta-learned model should use experience from training tasks to adapt to new tasks, although the quality of generalization depends on task diversity and the relationship between training and test tasks.
44. Which factor is important when designing a Meta-Learning benchmark?
- Ensuring that every task has an identical name
- Using only the training tasks for final evaluation
- Measuring only the size of the model file
- Creating appropriate task splits and evaluating adaptation on held-out tasks
Answer: D) Creating appropriate task splits and evaluating adaptation on held-out tasks
Explanation:
A good benchmark separates meta-training tasks from validation and test tasks. It should also define the support data available for adaptation, the query examples used for evaluation, and metrics appropriate to the target problem.
45. What is a potential disadvantage of using a very small support set during Meta-Learning?
- The available examples may not adequately represent the target task, leading to uncertain or inaccurate adaptation
- The model is guaranteed to achieve perfect accuracy
- The number of classes automatically becomes zero
- The model can no longer use pretrained parameters
Answer: A) The available examples may not adequately represent the target task, leading to uncertain or inaccurate adaptation
Explanation:
A small support set provides limited evidence about the target task. If its examples are unrepresentative or noisy, adaptation may produce poor predictions. Meta-Learning can improve sample efficiency, but it cannot guarantee reliable results in every low-data situation.
46. How can Meta-Learning be useful in natural language processing?
- By preventing language models from using contextual information
- By helping a model adapt to new text classification or language tasks with limited labeled examples
- By converting all sentences into identical strings
- By eliminating the need to define evaluation criteria
Answer: B) By helping a model adapt to new text classification or language tasks with limited labeled examples
Explanation:
Meta-Learning can support tasks such as intent classification, text categorization, and adaptation to new domains. Training across multiple related tasks can help a model learn an adaptation strategy that requires fewer labeled examples for a new task.
47. What is the main difference between metric-based and optimization-based Meta-Learning?
- Metric-based methods cannot use neural networks, while optimization-based methods always use databases
- Metric-based methods do not use examples, while optimization-based methods use only unlabeled data
- Metric-based methods classify using learned representations and similarity relationships, while optimization-based methods learn parameters or update procedures that support rapid adaptation
- Both approaches always use identical algorithms and objectives
Answer: C) Metric-based methods classify using learned representations and similarity relationships, while optimization-based methods learn parameters or update procedures that support rapid adaptation
Explanation:
Metric-based methods, such as Prototypical Networks, rely on distances or similarities in an embedding space. Optimization-based methods, such as MAML, learn initial parameters that can be adapted through task-specific optimization. Both approaches aim to generalize across tasks but use different mechanisms.
48. A research team wants an image classifier to recognize new animal species from only two labeled images per species. It has a collection of related training tasks and wants the model to learn how to adapt to new classes. Which approach is most appropriate?
- Train a separate model from scratch for every species without sharing knowledge
- Use only a fixed set of labels and prohibit new categories
- Evaluate the model only on the two support images for each species
- Use episodic Meta-Learning with few-shot tasks and evaluate on separate query images from held-out species
Answer: D) Use episodic Meta-Learning with few-shot tasks and evaluate on separate query images from held-out species
Explanation:
Episodic Meta-Learning can expose the model to many simulated few-shot classification tasks. Evaluating on held-out species with separate query examples measures whether the model can transfer its learned adaptation strategy to new categories rather than simply memorizing the support images.
49. A team uses MAML for a new classification task, but the model performs well on its support examples and poorly on query examples. Which issue should the team investigate?
- Overfitting during task-specific adaptation or a mismatch between meta-training and evaluation tasks
- The alphabetical order of the class names
- The number of comments in the source code
- The color of the visualization used to display results
Answer: A) Overfitting during task-specific adaptation or a mismatch between meta-training and evaluation tasks
Explanation:
Strong support-set performance with poor query-set performance can indicate overfitting during the inner loop. The team should examine adaptation steps, learning rates, support-set quality, task diversity, and the similarity between meta-training and evaluation tasks.
50. An AI team is developing a model that must adapt to new industrial defect categories using only a few labeled images per category. The team has multiple related training tasks, limited computing resources, and a requirement to measure performance on unseen categories. Which strategy is most appropriate?
- Train an independent large model from scratch for each defect category and skip held-out evaluation
- Use episodic Meta-Learning, compare a computationally practical method such as Prototypical Networks or Reptile with suitable baselines, and evaluate on held-out categories using separate query images
- Use the same images for meta-training and final testing to maximize reported accuracy
- Choose a method solely by its training accuracy without examining adaptation performance
Answer: B) Use episodic Meta-Learning, compare a computationally practical method such as Prototypical Networks or Reptile with suitable baselines, and evaluate on held-out categories using separate query images
Explanation:
Episodic Meta-Learning can train a model to adapt from a small support set. Prototypical Networks provide a metric-based option, while Reptile offers a first-order optimization-based approach. The team should compare methods against appropriate baselines, account for compute constraints, and evaluate on held-out defect categories using separate query examples to assess generalization.