Introduction to Artificial Intelligence Questions and Answers

Answer: A

The empirical finding that performance improves predictably with data and compute has driven the scale of modern systems and shifted the field's centre of gravity towards learning. It does not remove the need for architectural and algorithmic advances, which have continued in parallel. Storage cost and interpretability are separate matters, though interpretability generally worsens with scale. A reliable empirical relationship between resources and performance is unusual and it has shaped a decade of research strategy.

Answer: B

More data of the wrong distribution, or with noisy labels, or lacking the cases that matter most, adds little and can mislead, so curation frequently matters more than volume. Storage and processing are engineering constraints rather than fundamental limits, and correct additional data does not reduce accuracy. Recognising that data quality and coverage set the ceiling, rather than sheer quantity, redirects effort towards the part of the pipeline that actually binds.

Answer: C

A system trained on one population may perform badly on another that differs in ways the training data did not capture, and performance degrades silently rather than announcing itself. Monitoring deployed performance rather than trusting a test score is the standard response. Parameter drift, Bayesian updating and platform variation are separate concepts. A failure mode that produces confident wrong answers rather than obvious errors is particularly dangerous and deserves explicit monitoring.

Answer: D

A test set drawn from the same collection as the training data measures generalisation within that distribution and says nothing about performance on a population it does not represent. Repeated use of the same test set also inflates scores through selection. Hardware and sample size are lesser concerns. The distinction between generalising within a distribution and generalising to a new one is the most consequential subtlety in evaluating a learned system.

Answer: A

The first two are correct. Deep learning is a subset of machine learning, which is itself a subset of artificial intelligence, and large areas of the field involve search, planning and reasoning with no learning at all. Getting the nesting the wrong way round is a common error and it leads to the assumption that every problem in the field is a data problem, which misdirects a great deal of effort.

Answer: B

The four form a ladder, each adding a capability the previous one lacked: memory, then an explicit objective, then a graded preference over outcomes. Choosing the simplest architecture that the environment demands is the design principle, since each addition costs computation and design effort. Learning can be added to any of them, which is why it is treated as an orthogonal dimension rather than a fifth rung.

Answer: C

The first three are correct. Discrete refers to whether states, time and actions take finitely many distinct values, while the number of agents is the separate single agent or multi-agent dimension. The taxonomy is worth learning precisely because each dimension selects a different technique: memory for partial observability, probability for stochasticity, real time constraints for dynamism and game theory for multiple agents.

Answer: D

Language, vision and planning are all recognised subfields with their own conferences and literature. Numerical analysis is a branch of mathematics concerned with the accuracy and stability of computation, used throughout the field as a tool but not part of it. Distinguishing the disciplines a subject draws on from the subject itself keeps its boundaries clear, which matters when assessing whether a claimed result belongs to it at all.

Answer: A

Both are true and the second is the reason for the first. A system that plays chess at superhuman level cannot play draughts, hold a conversation or fold a towel, and must be rebuilt or retrained for each new task. Whether generality will emerge from scaling current methods or requires something new is genuinely open. Being explicit about narrowness is the most useful discipline when reading claims about the field's progress.

Answer: B

The second is correct and the first is exactly what the test was designed to avoid. Turing deliberately replaced questions about mechanism with a purely behavioural criterion, because he regarded the mechanistic question as unanswerable. That choice is both the test's strength, since it can actually be conducted, and its weakness, since behaviour can be imitated without the capabilities the test was meant to detect. The pair captures the whole debate about it in two sentences.