Selected Topics Reviewer

Why does semantic search improve retrieval quality in RAG systems?

It retrieves documents with similar meaning using embeddings

What is the primary difference between discriminative and generative machine learning models?

Discriminative models learn decision boundaries to classify, while generative models learn probability distributions to generate new data

Why do standard RNNs struggle with predicting the next word in very long sequences.

Because RNNs have difficulty retaining information over long time steps due to vanishing gradients.

In RAG pipelines, query rewriting using an LLM primarily improves retrieval quality by:

Reducing ambiguity and aligning queries with document vocabulary

What does tSNE primarily accomplish in data analysis?

Visualizing high dimensionality data.

Large models reach the same test loss using fewer tokens, demonstrating higher sample efficiency.

True

What are the two main components that make up a Retrieval Augmented Generation (RAG) system?

Retriever and generator

N-gram language models can effectively handle long-range dependencies in text because they consider the entire context of a sentence.

false

Graph RAG retrieves connected entities and relationships rather than just similar text chunks.

True

How does RAG concretely reduce hallucinations?

By grounding responses in retrieved evidence

RAG is a hybrid model that combines only a retrieval component without any generation capabilities.

False

Which of the following measures best ensures reliability and fairness when deploying AI systems capable of autonomous decision-making?

Implementing monitoring, human-in-the-loop review, and periodic auditing of AI outputs.

Why do Transformers handle long-range dependencies better than RNNs and LSTMs?

false

How does RAG differ from traditional LLMs in terms of knowledge source?

RAG uses dynamic external information while traditional LLMs rely on static training data

What is a soft prompt in the context of LLMs ?

Learned embeddings added to the model input without changing the model's weights.

Which of the following tasks can most naturally be suited for a Autoencoder (AE) and Variational Autoencoder (VAE) rather than a GAN or diffusion model?

Learning a compact latent representation of input data for anomaly detection.

According to the compute-efficient scaling law, what is the primary reason training small models to full convergence is considered inefficient?

Small models cannot fully utilize the available compute, wasting computational resources.

During preprocessing, why might you choose to convert all text to lower case before tokenization and stopword removal?

To ensure that different capitalizations of the same word are treated consistently and mapped to the same token.

What is the purpose of positional encoding in transformer architecture?

To preserve information about word order in the sequence

Which of the following is NOT mentioned as a benefit of Explainable AI?

Reducing the computational cost of machine learning model training

What is the key idea behind semi-supervised learning?

Unlabeled data (in addition to limited labeled data) helps reveal the underlying structure or distribution of the data

What is the key idea behind the Query-by-Committee active learning strategy?

The model maintains multiple models and queries the instance for which these models disagree the most

What are the key benefits of multi-task learning?

Faster training, improved performance per task, shared representation and data augmentation

Expected model change is an active learning strategy that selects the data points that would cause the largest change in the model parameters if added to the training set.

True

In multi-task learning, a single model is trained on multiple tasks simultaneously.

true

Which of the following best explains why multi-task learning can be seen as a form of implicit data augmentation?

Training on multiple tasks exposes the model to a broader variety of signals and patterns

Reinforcement learning is a type of supervised learning.

False

A researcher trains a model on a large labeled dataset of chest X-

rays from Hospital A and then deploys it on data from Hospital B, where imaging devices and patient demographics differ significantly. The task (disease classification) remains the same, but performance drops due to distribution shift. Which approach is most appropriate to address this issue?

Domain adaptation

Consider training a DANN where the domain classifier achieves very high accuracy in distinguishing source from target samples. What does this indicate about the learned feature representation?

The feature extractor has failed to learn domain

What is the main risk associated with the self-training algorithm in semi-self-training algorithm in semi-supervised learning?

Incorrect pseudo-labels (predicted labels) may propagate errors

What is the main purpose of Explainable AI (XAI)?

To provide insights into how AI models function and enable more accountable, ethical, and fair decision

In hard parameter sharing for Multi-Task Learning, what is the main trade-off compared to soft parameter sharing?

Hard sharing reduces memory usage but may limit task-specific specialization

What is the main idea behind transfer learning?

Pre-training a model on a large dataset and then fine-tuning it on a new smaller dataset.

In the example of the autonomous drone, how does the reinforcement learning algorithm differ from a supervised learning approach?

reinforcement maximizes long term rewards, requires an agent to explore an environment through trial and error, does not require correction action for every task

What is the main difference between passive learning and active learning?

Passive learning selects data points randomly for labeling, while active learning selects (for labeling) the most informative unlabeled data points

A model is pretrained on ImageNet and then fine-tuned on a small medical imaging dataset. Under which condition is freezing the backbone most appropriate?

When the target dataset is small and similar low-level features are expected to transfer

In a label propagation algorithm (for semi-supervised learning), consider an unlabeled node connected to 5 neighbors: 3 from Class A with very weak similarity weights, and 2 from Class B with a very strong similarity weight. What is the most likely outcome after several propagation steps?

The node will be classified as Class B due to stronger similarity influence

Why might Expected Model Change select a sample that uncertainty sampling does not?

Because it considers how much the model parameters would change after training with that sample

What is the key idea behind the Query-by-Committee active learning strategy?

The model maintains multiple models and queries the instance for which these models disagree the most

Which of these models is considered to have the highest level of explainability?

Linear Regression

Consider a 3-class classifier with the following predicted probabilities for two unlabeled samples: x1: [0.40, 0.35, 0.25] x2: [0.45, 0.30, 0.25] Which statement is correct in the context of Active Learning ?

Both least confident and margin sampling will select x1

In zero-shot learning, a model can recognize classes it has never seen during training.

true

In the context of Meta-Learning, what is the fundamental difference between the "meta-training" phase and the "meta-test" phase?

Meta-training uses a set of training tasks to learn how to adapt quickly and effectively, while meta-test evaluates the model's ability to learn a new task efficiently using only a few labeled examples.

Why might full fine-tuning outperform freezing or partial fine-tuning when adapting to a new task?

It allows all layers to adapt to potentially different feature distributions

Deep learning models have very high explainability compared to other AI models.

False

In hard parameter sharing for Multi-Task Learning, what is the main trade-off compared to soft parameter sharing?

Hard sharing reduces memory usage but may limit task-specific specialization

In the learning algorithm described for DQN, we repeatedly create an artificial training set to which we apply supervised learning, where the input is x = (s, a) and the target, constructed using Bellman’s equations, is y = ......... ?

y = R(s) + γ max_{a'} Q(s', a') where s' is the state reached after taking action a in state s

A policy in reinforcement learning is a function that maps states to actions.

True

Consider two human activity recognition tasks: T1: Activities (walking, sitting, standing) from subject s1 in environment e1 T2: Activities (running, jumping, cycling) from subject s2 in environment e2 Which of the following best explains why these are considered different tasks?

Both p(x) and p(y|x) may differ

According to the formal definition presented in the lecture, what elements constitute a machine learning "task" Ti ?

An input data space distribution Pi(x), a labeling distribution Pi(y|x), and an objective function Li

Which AI model is considered to have the lowest level of explainability?

Deep Neural Networks

Semantic search in RAG looks for documents containing the exact words found in the user's prompt.

False

Which of the following best describes the BLEU metric?

It measures average precision across multiple n

What are the two main components that make up a Retrieval

Augmented Generation (RAG) system?

According to the scaling laws research, which factor has the strongest impact on reducing test loss in neural language models?

Dataset size (measured in tokens)

Which observation best explains why transfer learning became effective only after large

scale datasets like ImageNet were introduced?

A company deploys an LLM-based assistant for internal policy questions. The LLM frequently gives confident but incorrect answers about recently updated policies. Which limitation of traditional LLMs is RAG primarily designed to address in this scenario?

Dependence on static, historical training data

A standard autoencoder has a small bottleneck layer (i.e., constrained to have very few neurons in the middle layer). Why does this bottleneck help the network learn useful features instead of just copying the input?

Because the bottleneck constraint forces the network to compress information and retain only important features.

Consider the example of an autonomous drone in the context of Reinforcement Learning. What information does the drone's state consist of?

The drone's position and motion information.

Identify the type of fairness metric that measures the difference in true positive rates between different groups.

Equal Opportunity Difference (EOD)

Query-by-committee is an active learning strategy that maintains multiple models and selects the data points for which the models disagree the most.

True

A loan applicant is rejected by a ML model. The system outputs: "Your loan would have been approved if your income were $10,000 higher." This is an example of:

Counterfactual explanation