Introduction to Artificial Intelligence Questions and Answers

Answer: C

A system trained on historical data reproduces the patterns in it, including discriminatory ones, and can amplify them, so its outputs differ systematically across groups in ways that are unjust. It is a social harm arising from a technical process. Estimator bias, inductive bias and search order are all legitimate technical uses of the word bias and none of them is this. Distinguishing the technical senses from the social one prevents a great deal of confused discussion.

Answer: D

Learned systems reproduce whatever regularities exist in their training data, so historical under-representation or discriminatory decisions are learned as patterns. Choice of objective and of features contributes as well. The algorithm's mathematics is neutral between groups, hardware is irrelevant, and deliberate discrimination is rare and a different problem. Locating the origin of a harm correctly determines whether the remedy is better data, a different objective or a constraint on the output.

Answer: A

Without assumptions no generalisation from finite data is possible at all, so every learning algorithm embodies a preference, such as for smooth functions or for simple hypotheses. The choice of that bias is what makes one algorithm suit a problem better than another. Estimator bias, fairness and the generalisation gap are separate concepts sharing vocabulary. Recognising that assumption free learning is impossible is the content of the no free lunch results.

Answer: B

Averaged over every possible problem, all algorithms perform equally, so any algorithm's advantage comes from its assumptions matching the structure of the problems actually encountered. That is why choosing a method is choosing a set of assumptions rather than finding a universally best tool. Learning from finite data is certainly possible given assumptions, and computation and interpretability are separate matters. A theorem that forbids a universal solution redirects effort towards understanding the problem instead.

Answer: C

The Turing machine establishes the class of computable functions, and with the Church Turing thesis it bounds what any physical computer can do, which frames the question of whether intelligence is computable. It is a theoretical construct rather than a physical machine, shares nothing architecturally with neural networks, and makes no claim about brains. Knowing what lies outside computation altogether is a useful boundary when assessing claims about what machines might eventually do.

Answer: D

The thesis identifies the informal notion of an effective procedure with the formal class of Turing computable functions, and it is a claim about definitions rather than a theorem. It does not assert that every function is computable, since most are not, nor does it make any claim about human thought or about the solubility of particular problems. Distinguishing a definitional thesis from an empirical claim is important when it is invoked in arguments about machine intelligence, as it frequently is.

Answer: A

Newell and Simon proposed that a system capable of manipulating symbol structures has everything needed for general intelligent action, which is the founding claim of the symbolic tradition. It is an empirical hypothesis rather than a proven result, and embodied and connectionist researchers dispute it. Storage medium, embodiment and symbol grounding are related questions rather than the hypothesis itself. Identifying the founding assumption of a research programme clarifies what its critics are actually attacking.

Answer: B

A symbol manipulated purely syntactically has no connection to what it is supposed to refer to unless something ties it to perception and action, and providing that connection is the grounding problem. It underlies the Chinese room argument and motivates embodied approaches. Storage, translation and uniqueness are engineering matters. A question about how meaning enters a formal system at all is philosophical in origin and has practical consequences for how systems are built.

Answer: C

Reasoning and calculation, which people find effortful, proved comparatively easy to automate, while perception and movement, which any child performs effortlessly, proved extremely hard. The suggested explanation is that the latter have been optimised by evolution for far longer. The other options describe no recognised paradox. An observation that inverts the intuitive ordering of difficulty is a useful corrective when estimating how hard a new task will be.

Answer: D

Interpreting a scene depends on a vast body of unstated expectations about objects, lighting, occlusion and physics that people acquire without instruction, whereas the rules of logic are few and explicit. That is the same commonsense knowledge difficulty in another guise. Arithmetic volume and sensor quality are secondary, and both areas have long histories. A task whose difficulty lies in unstated knowledge rather than in computation resists exactly the improvements that hardware provides.