D429 Introduction to AI for Computer Scientists - Set 1 - Part 2

Correct Answer: Q-learning reinforcement approach

Option b is correct because Q-learning handles sparse rewards effectively, learning optimal action values for personalizing content difficulty from periodic feedback.
Option a is incorrect because supervised batch learning requires complete labeled datasets rather than adapting to sparse, intermittent rewards from student performance.
Option c is incorrect because unsupervised clustering groups similar students but cannot learn optimal content sequencing from performance-based rewards effectively.
Option d is incorrect because regression models predict continuous values but don't learn sequential decision-making policies from sparse reward signals.

Correct Answer: Algorithmic transparency accountability

Option b is correct because transparency and accountability ensure content moderation decisions can be understood, appealed, and corrected when necessary.
Option a is incorrect because focusing solely on efficiency might lead to over-censorship or under-moderation without considering fairness and user rights.
Option c is incorrect because complete automation without human oversight risks systematic errors and bias in content moderation decisions affecting expression.
Option d is incorrect because minimal human oversight could allow algorithmic biases to persist unchecked, potentially harming legitimate expression rights.

Correct Answer: Universal transformer architectures

Option b is correct because transformer architectures with self-attention can learn universal patterns that capture meaning across different languages effectively.
Option a is incorrect because language-specific tokenizers alone cannot bridge semantic understanding across different languages without shared representational learning capabilities.
Option c is incorrect because monolingual word lists lack the cross-linguistic semantic relationships needed for multilingual understanding and context transfer.
Option d is incorrect because single language models cannot share knowledge or transfer understanding between different languages in the system.

Correct Answer: Diverse data collection

Option b is correct because collecting diverse data from multiple geographic regions ensures the model learns patterns that generalize across populations.
Option a is incorrect because increasing training epochs on biased data will likely lead to overfitting rather than improving geographic generalization.
Option c is incorrect because reducing complexity might hurt performance without addressing the fundamental lack of geographic diversity in training data.
Option d is incorrect because static feature selection doesn't address the core issue of training data lacking geographic diversity and representation.

Correct Answer: Partially observable dynamic

Option b is correct because forest navigation involves limited sensor visibility through trees and dynamic changes from falling branches and moving animals.
Option a is incorrect because forests aren't fully observable due to visual obstructions, and the environment changes with falling trees and moving animals.
Option c is incorrect because animal movements and falling trees make the environment non-deterministic, and the delivery task is sequential, not episodic.
Option d is incorrect because forest navigation involves continuous spatial movement and unknown elements like animal behavior patterns and tree conditions.

Correct Answer: Adversarial training methods

Option b is correct because adversarial training explicitly includes manipulated examples during training, making the model robust against deliberate deception attempts.
Option a is incorrect because standard supervised training doesn't prepare the model for adversarial attacks designed to fool the recognition system.
Option c is incorrect because unsupervised pretraining alone doesn't provide specific defenses against adversarial examples targeting the security system.
Option d is incorrect because transfer learning transfers knowledge but doesn't inherently provide robustness against adversarial attacks on the system.

Correct Answer: Hybrid filtering approach

Option b is correct because hybrid filtering combines content-based and collaborative methods, providing initial recommendations while adapting to preferences.
Option a is incorrect because random selection provides poor user experience and doesn't leverage available information about items or similar users.
Option c is incorrect because historical averaging ignores individual differences and cannot personalize recommendations for new users with unique preferences.
Option d is incorrect because static popularity rankings don't personalize recommendations or adapt as the system learns about new user preferences.

Correct Answer: Human oversight integration

Option b is correct because maintaining human oversight ensures the AI system remains under control and prevents autonomous harmful decision-making patterns.
Option a is incorrect because maximum autonomy increases the risk of the gorilla problem where the AI becomes uncontrollable and potentially harmful.
Option c is incorrect because complete automation removes human control, directly contributing to the gorilla problem of uncontrollable AI systems.
Option d is incorrect because minimal constraints allow the AI too much freedom, potentially leading to the uncontrollable behaviors the gorilla problem describes.

Correct Answer: Semantic network graphs

Option b is correct because semantic networks provide visual, traceable paths showing relationships and reasoning that auditors can follow and verify.
Option a is incorrect because deep neural networks often lack interpretability, making it difficult to provide clear explanations required by financial regulations.
Option c is incorrect because black box models by definition cannot provide the transparent, auditable reasoning paths required for regulatory compliance.
Option d is incorrect because unstructured text logs lack the organized, traceable structure needed for systematic audit and regulatory review processes.

Correct Answer: LSTM memory cells

Option b is correct because LSTM memory cells with gates can selectively remember and forget information, maintaining relevant context across topic transitions.
Option a is incorrect because stateless designs cannot maintain any conversation history or context, making coherent multi-topic conversations impossible to manage.
Option c is incorrect because bag-of-words accumulation loses sequential information and cannot track how topics transition and relate during conversations.
Option d is incorrect because one-hot encoding represents states as binary vectors without the ability to maintain nuanced conversational context information.

Correct Answer: Binary classification problem

Option b is correct because predicting dropout risk is naturally a binary classification determining whether students will drop out or continue.
Option a is incorrect because multi-class classification unnecessarily complicates the problem when the outcome is simply dropout versus continuation in the program.
Option c is incorrect because regression predicts continuous values rather than the categorical outcome of whether a student drops out or continues.
Option d is incorrect because unsupervised clustering might group similar students but cannot directly predict the specific dropout risk outcome needed.

Correct Answer: Learning agent architecture

Option b is correct because learning agents observe user behavior patterns, adapt their responses, and improve comfort settings over time automatically.
Option a is incorrect because simple reflex agents only respond to current conditions without learning from patterns of user preferences over time.
Option c is incorrect because static rule systems cannot adapt or learn new preferences as user behaviors and comfort requirements change seasonally.
Option d is incorrect because hardcoded responses cannot adjust to individual user preferences or learn from observed behavior patterns in the home.

Correct Answer: Feature fusion techniques

Option b is correct because feature fusion combines different data types into unified representations that capture sentiment from multiple information sources.
Option a is incorrect because processing only text reviews ignores valuable sentiment signals from star ratings and purchase behavior patterns.
Option c is incorrect because numerical scaling alone cannot integrate text reviews with numerical data for comprehensive sentiment analysis effectively.
Option d is incorrect because independent analysis streams fail to capture the interactions and correlations between different types of customer feedback.

Correct Answer: Conditional probability relationships

Option b is correct because conditional probability captures how disease likelihood changes given specific symptoms, essential for medical diagnostic reasoning.
Option a is incorrect because joint probability distribution shows all combinations but doesn't specifically model how symptoms influence disease probability diagnostically.
Option c is incorrect because marginal independence incorrectly assumes symptoms don't relate to diseases, which contradicts medical diagnostic knowledge fundamentally.
Option d is incorrect because deterministic values cannot represent the inherent uncertainty in medical diagnosis where symptoms suggest but don't guarantee diseases.

Correct Answer: Cooperative multiagent environment

Option b is correct because warehouse robots must cooperate to avoid collisions, share workspace, and collectively optimize package movement efficiency.
Option a is incorrect because multiple robots working together clearly constitute a multiagent system rather than a single agent environment.
Option c is incorrect because warehouse robots need cooperation rather than competition to achieve the common goal of efficient package handling.
Option d is incorrect because isolated independent agents would cause collisions and inefficiencies without coordination in the shared warehouse space.

Correct Answer: Data veracity accuracy

Option b is correct because veracity ensures data truthfully represents reality without biases that could lead to discriminatory AI decisions.
Option a is incorrect because processing speed doesn't address the fundamental bias issues that cause discriminatory outcomes in AI decisions.
Option c is incorrect because storage optimization is unrelated to addressing the ethical concerns of bias and discrimination in training data.
Option d is incorrect because format uniformity standardizes structure but doesn't eliminate biases embedded in the actual data content values.

Correct Answer: Semantic analysis rules

Option b is correct because semantic rules analyze meaning and context to distinguish factual claims from subjective opinion expressions in language.
Option a is incorrect because tokenization only breaks text into pieces without understanding the semantic difference between facts and opinions expressed.
Option c is incorrect because character encoding converts text to computer-readable format but doesn't analyze meaning or distinguish statement types.
Option d is incorrect because frequency counting tracks word occurrence but cannot determine whether statements express facts or personal opinions effectively.

Correct Answer: Dynamic Bayesian networks

Option b is correct because dynamic Bayesian networks model temporal dependencies and update probabilities as conditions like weather and traffic change.
Option a is incorrect because static probability tables cannot adapt to changing conditions or model temporal dependencies in supply chain dynamics.
Option c is incorrect because simple probability rules lack the sophistication to model complex interactions between multiple time-varying factors affecting deliveries.
Option d is incorrect because fixed distribution models cannot capture the evolving nature of weather, traffic, and supplier reliability over time.

Correct Answer: Differential privacy implementation

Option b is correct because differential privacy protects individual privacy while still allowing aggregate analysis for legitimate security purposes effectively.
Option a is incorrect because maximum surveillance coverage prioritizes security over privacy, failing to balance these competing ethical concerns appropriately.
Option c is incorrect because unrestricted data collection violates privacy principles and likely breaches data protection regulations requiring consent and limitations.
Option d is incorrect because minimal oversight could lead to privacy violations and abuse without proper checks on surveillance system usage.

Correct Answer: Restricted output filtering

Option b is correct because output filtering can identify and prevent generation of content that could enable malicious activities or harm.
Option a is incorrect because maximum creative freedom without restrictions could allow generation of harmful content like malware or dangerous instructions.
Option c is incorrect because unconstrained generation explicitly avoids the safety measures needed to prevent creation of potentially harmful content outputs.
Option d is incorrect because prioritizing speed over safety could result in harmful content being generated before safety checks can intervene.