I85PREP
⌘ K
HIRAF study materials
HIRAF study resources

HIRAF practice questions

One question at a time. Build your HIRAF understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1Fundamentals of Artificial Intelligence

Which term describes the field of computer science concerned with building systems that perform tasks normally requiring human intelligence?

  1. Relational database design
  2. Compiler construction
  3. Digital signal processing
  4. Artificial intelligence Correct answer.

Explanation

Artificial intelligence is defined by its goal of reproducing or simulating intelligent behaviour such as reasoning, learning, and perception in machines. The other disciplines listed focus on signal transformation, structured data storage, and language translation into machine code, none of which is defined by intelligence.

Question 2Fundamentals of Artificial Intelligence

In the standard agent–environment formulation, what is the role of a percept?

  1. The rule that maps states to rewards
  2. The measure of how optimal a policy is
  3. The agent's perceptual input at a given moment Correct answer.
  4. The set of all actions available to the agent

Explanation

A percept is the agent's individual sensory reading at one instant, and the percept sequence is the complete history of such readings. Actions, rewards, and performance measures are separate components of the agent-environment framework, so none of them names the perceptual input itself.

Question 3Fundamentals of Artificial Intelligence

A search algorithm expands nodes in order of increasing value of f(n) = g(n) + h(n). Which algorithm is this?

  1. Depth-first search
  2. Breadth-first search
  3. A* search Correct answer.
  4. Uniform-cost search

Explanation

A* ranks nodes by the sum of the path cost so far, g(n), and the heuristic estimate of remaining cost, h(n). Uniform-cost search uses only g(n), while breadth-first and depth-first search order nodes by structure rather than by an estimated total cost.

Question 4Fundamentals of Artificial Intelligence

Which statement best characterises supervised learning?

  1. The model learns from labelled input-output pairs Correct answer.
  2. The model improves by trial and error using rewards
  3. The model clusters unlabelled data by similarity
  4. The model memorises the training set without generalising

Explanation

Supervised learning fits a mapping from inputs to outputs using examples where the target answer is provided. Clustering is unsupervised, reward-driven trial and error is reinforcement learning, and pure memorisation is a failure mode rather than a learning paradigm.

Question 5Fundamentals of Artificial Intelligence

What does the bias term contribute in an artificial neuron computing a weighted sum of its inputs?

  1. It guarantees the neuron's output is always positive
  2. It normalises each input to zero mean
  3. It shifts the activation threshold independently of the inputs Correct answer.
  4. It removes the need for a learning rate

Explanation

The bias is an extra weight applied to a constant input, which shifts the decision boundary or activation threshold away from the origin. It does not rescale the inputs, force positivity, or replace hyperparameters such as the learning rate.

Question 6Fundamentals of Artificial Intelligence

In propositional logic, what is a tautology?

  1. A formula false under every possible assignment of truth values
  2. A formula whose truth depends on the assignment
  3. A formula true under every possible assignment of truth values Correct answer.
  4. A formula containing no logical connectives

Explanation

A tautology is valid: it evaluates to true in every row of its truth table. Formulas that are always false are contradictions, and formulas whose value varies with the assignment are contingent.

Question 7Fundamentals of Artificial Intelligence

Which of the following is a classic example of a combinatorial optimisation problem often used to illustrate heuristic search?

  1. Computing a square root
  2. Sorting a list of integers
  3. Parsing a regular expression
  4. The travelling salesman problem Correct answer.

Explanation

The travelling salesman problem asks for the shortest tour visiting every city once and is NP-hard, so heuristics such as nearest neighbour or local search are commonly applied. Sorting, root extraction, and regular expression parsing all have efficient deterministic procedures.

Question 8Fundamentals of Artificial Intelligence

What distinguishes strong AI from weak AI as these terms are traditionally used?

  1. Strong AI uses symbolic methods; weak AI uses statistical methods
  2. Strong AI runs on more powerful hardware than weak AI
  3. Strong AI is deployed commercially; weak AI remains in laboratories
  4. Strong AI claims genuine understanding or consciousness; weak AI only simulates intelligent behaviour Correct answer.

Explanation

The strong-weak distinction concerns the nature of machine intelligence, not its implementation or hardware. Weak AI treats the machine as a simulation of cognition, whereas strong AI asserts that an appropriately programmed machine can literally possess a mind.

Question 9Fundamentals of Artificial Intelligence

A heuristic h is admissible for A* search. What property does this guarantee?

  1. h never overestimates the true cost to reach the goal Correct answer.
  2. h is independent of the problem domain
  3. h is monotonically decreasing along any path
  4. h is always exactly equal to the true remaining cost

Explanation

Admissibility means h(n) is a lower bound on the true cost of the cheapest path from n to a goal. With a non-negative step cost this ensures A* with a graph search that avoids reopening can still return an optimal solution. Equality with the true cost is stronger than admissibility, and monotonicity is a separate condition called consistency.

Question 10Fundamentals of Artificial Intelligence

In a minimax search of a game tree, what does alpha-beta pruning accomplish?

  1. It changes the evaluation function to favour the maximising player
  2. It returns the same root value while skipping branches that cannot affect the result Correct answer.
  3. It removes the need for a depth limit in the search
  4. It returns an approximate value that may differ from full minimax

Explanation

Alpha-beta pruning keeps track of bounds on the best values found so far and cuts off branches that provably cannot influence the final decision. The root value is identical to plain minimax, so the benefit is speed rather than a change in the result. It does not alter evaluation functions or depth limits.

Question 11Fundamentals of Artificial Intelligence

A Bayesian network encodes a joint probability distribution over variables. What does the absence of an edge between two nodes typically represent?

  1. A conditional independence assumption given the node's parents Correct answer.
  2. That the two variables are statistically independent unconditionally
  3. That the variables have equal prior probabilities
  4. That no data were collected for the pair of variables

Explanation

Each node's distribution is conditioned only on its parents, so missing edges encode conditional independences relative to the parent set. Unconditional independence does not follow in general, because dependence can flow through other nodes. Missing edges say nothing about data availability or equal priors.

Question 12Fundamentals of Artificial Intelligence

Which situation is an example of overfitting in a trained classifier?

  1. High accuracy on training data but much lower accuracy on held-out data Correct answer.
  2. Accuracy that improves steadily as more training data are added
  3. Equal accuracy on training and held-out data
  4. Low accuracy on both training and held-out data

Explanation

Overfitting occurs when a model captures noise specific to the training set, producing a large gap between training and test performance. Low accuracy on both sets indicates underfitting instead, and stable or improving generalisation shows no overfitting.

Question 13Fundamentals of Artificial Intelligence

A system must decide which of several actions to take to maximise expected long-term reward while the environment's transition probabilities are unknown. Which framework is designed for this setting?

  1. Dimensionality reduction
  2. Constraint satisfaction
  3. Supervised classification
  4. Reinforcement learning Correct answer.

Explanation

Reinforcement learning addresses sequential decision making where an agent learns from interaction and reward signals without a model of the environment. Supervised classification needs labelled targets, dimensionality reduction compresses features, and constraint satisfaction seeks assignments meeting fixed constraints.

Question 14Fundamentals of Artificial Intelligence

What is the purpose of the resolution inference rule in automated reasoning?

  1. To assign truth values to all atoms simultaneously
  2. To derive new clauses by eliminating a complementary literal pair Correct answer.
  3. To check whether a formula is satisfiable by enumeration
  4. To convert arbitrary formulas into disjunctive normal form

Explanation

Resolution takes two clauses containing a literal and its negation and produces a resolvent without that pair, which is a sound and refutation-complete rule for clause form. It does not perform normal-form conversion, truth assignment, or brute-force enumeration.

Question 15Fundamentals of Artificial Intelligence

In the standard genetic algorithm cycle, which operation recombines parts of two parent solutions?

  1. Crossover Correct answer.
  2. Mutation
  3. Selection
  4. Fitness evaluation

Explanation

Crossover combines genetic material from two parents to create offspring that mix parental traits. Selection chooses which individuals reproduce, mutation randomly alters genes, and fitness evaluation scores candidates; none of these recombines two parents.

Question 16Fundamentals of Artificial Intelligence

A knowledge base contains the rule 'if it rains then the ground is wet' and the fact 'it rains'. Using modus ponens, what can be concluded?

  1. The ground is wet Correct answer.
  2. Nothing can be concluded
  3. It does not rain
  4. The ground is dry

Explanation

Modus ponens permits deriving the consequent of an implication once its antecedent is established. Given the rule and the fact that the antecedent holds, the consequent follows deductively. Denying the antecedent or negating the conclusion would be invalid inferences.

Question 17Fundamentals of Artificial Intelligence

What does the term 'frame problem' refer to in knowledge representation?

  1. The impossibility of representing any real-world object
  2. The difficulty of specifying which facts stay unchanged after an action Correct answer.
  3. The problem of rendering graphical interfaces
  4. The limited memory available for storing frames

Explanation

The frame problem concerns how to represent the many facts that persist through an action without writing explicit axioms for every unaffected proposition. It is a representational and computational issue in action reasoning, unrelated to memory limits, user interfaces, or representability in principle.

Question 18Fundamentals of Artificial Intelligence

A dataset contains 90 positive and 10 negative examples. A classifier predicts positive for every instance. What accuracy does it achieve, and why can accuracy mislead here?

  1. 100 percent, because all positives are correct
  2. 90 percent, because the majority class dominates the metric Correct answer.
  3. 50 percent, because the classes are balanced
  4. 10 percent, because negatives are counted only

Explanation

Predicting the majority class for all instances yields exactly the majority proportion, which is 90 of 100 here. Accuracy counts correct predictions regardless of class, so a trivial majority-class rule can appear strong while detecting no negatives at all. Metrics such as recall or balanced accuracy expose this failure.

Question 19Fundamentals of Artificial Intelligence

In first-order logic, what is the difference between a predicate and a function symbol?

  1. A predicate returns a truth value; a function returns an object Correct answer.
  2. Both return objects but differ in name length
  3. A predicate returns an object; a function returns a truth value
  4. Both return truth values but differ in arity

Explanation

Predicates form atomic formulas that are true or false of their arguments, whereas function symbols build terms that denote objects in the domain. This distinction is what allows quantified statements to refer to objects and assert properties about them. Arity and naming conventions do not define the difference.

Question 20Fundamentals of Artificial Intelligence

Which statement about unsupervised and semi-supervised learning is accurate?

  1. Unsupervised learning always produces the same clusters regardless of initialisation
  2. Semi-supervised learning uses a small labelled set together with a larger unlabelled set Correct answer.
  3. Unsupervised learning requires labelled targets for every example
  4. Semi-supervised learning discards all unlabelled data

Explanation

Semi-supervised learning exploits abundant unlabelled data alongside limited labels to improve generalisation. Unsupervised learning by definition needs no labels, and clustering results such as k-means can depend on initialisation and hyperparameters.

Question 21Fundamentals of Artificial Intelligence

What is the principal advantage of an informed search strategy over an uninformed one?

  1. It always finds a shorter path than any uninformed method
  2. It guarantees a solution in constant time
  3. It removes the need for a goal test
  4. It uses domain knowledge to focus expansion and often expands fewer nodes Correct answer.

Explanation

Informed search uses a heuristic estimate of distance to the goal to prioritise promising nodes, typically reducing the number of expansions. It does not guarantee constant time, cannot dispense with the goal test, and optimality depends on heuristic properties rather than being automatic.

Question 22Fundamentals of Artificial Intelligence

A planning problem is described by states, actions with preconditions and effects, an initial state, and a goal. Which statement about such formulations is correct?

  1. Preconditions describe the resulting state; effects determine applicability
  2. Actions are applicable regardless of the current state
  3. Both preconditions and effects describe only the goal test
  4. Preconditions determine when an action is applicable; effects describe the resulting state Correct answer.

Explanation

In classical planning an action may be applied only when its preconditions hold in the current state, and applying it changes the state according to its effects. This separation is what allows planners to reason about applicability and change. The goal is tested against states, not against action descriptions.

Question 23Fundamentals of Artificial Intelligence

In a feedforward network, what is the effect of using a linear activation function in every hidden layer?

  1. The network gains greater expressive power than with nonlinear units
  2. The network no longer requires training data
  3. The network collapses to an equivalent single linear transformation Correct answer.
  4. The network can represent any Boolean function exactly

Explanation

A composition of linear functions is itself linear, so stacking linear hidden layers adds no representational power beyond a single affine map. Nonlinear activations are what allow multilayer networks to approximate complex functions. Training data requirements are unaffected by the choice of activation.

Question 24Fundamentals of Artificial Intelligence

Consider a search space where every step cost is 1. Which pair of algorithms is guaranteed to return a minimum-cost solution?

  1. Greedy best-first search and depth-limited search
  2. Depth-first search and greedy best-first search
  3. Breadth-first search and iterative deepening depth-first search Correct answer.
  4. Hill climbing and simulated annealing

Explanation

With uniform step costs, breadth-first search finds the shallowest goal, and iterative deepening reproduces breadth-first order while using linear memory. Depth-first search can return a deeper, costlier goal, greedy best-first search ignores path cost, and local search methods offer no completeness or optimality guarantee.

Question 25Fundamentals of Artificial Intelligence

A propositional knowledge base entails a sentence if the sentence is true in every model of the knowledge base. Which statement follows from this definition?

  1. Entailment requires the sentence to appear syntactically in the knowledge base
  2. Entailment depends on the order in which sentences were added
  3. Entailment can be checked by testing whether the knowledge base conjoined with the negation is unsatisfiable Correct answer.
  4. Entailment holds whenever the knowledge base and sentence share the same atoms

Explanation

By the deduction and refutation principles, the knowledge base entails the sentence exactly when no model satisfies the knowledge base while falsifying the sentence, that is, when the conjunction with the negation is unsatisfiable. Entailment is semantic rather than syntactic, so literal presence, shared vocabulary, and insertion order are irrelevant.

Question 26Fundamentals of Artificial Intelligence

Which definition best describes artificial intelligence as an academic field?

  1. The study of how to manufacture faster computer hardware for data centres
  2. The study of building systems that perceive, reason, and act in ways we would call intelligent Correct answer.
  3. The study of cataloguing every fact that humans currently know
  4. The study of writing programs that follow fixed scripts without adaptation

Explanation

Artificial intelligence is the science of constructing agents that take in information from an environment and choose actions aimed at achieving goals. Reasoning, learning, perception, and planning all fall inside this scope, while hardware manufacturing and static scripting do not. Encyclopaedic fact storage is a database concern rather than intelligence.

Question 27Fundamentals of Artificial Intelligence

An agent that selects actions using a fixed lookup table from every possible percept sequence to an action is called what kind of agent?

  1. A learning agent that improves from experience
  2. A utility-based agent that ranks outcomes
  3. A simple reflex agent implemented as a table Correct answer.
  4. A goal-based agent that plans ahead

Explanation

A simple reflex agent maps the current percept directly to an action through condition-action rules, and the table-driven version enumerates every percept sequence explicitly. It performs no lookahead, keeps no model of future consequences, and does not revise its behaviour from experience. Planning and utility comparison belong to richer agent designs.

Question 28Fundamentals of Artificial Intelligence

A search tree has branching factor 2 and uniform step cost 1. Breadth-first search is run to depth 3. How many nodes does it expand, counting the root as depth 0?

  1. 15 Correct answer.
  2. 8
  3. 7
  4. 14

Explanation

Breadth-first search expands every node at depth 0, 1, 2, and 3 before stopping once a goal at depth 3 is generated. The counts are 1 + 2 + 4 + 8, giving 15 expanded nodes in the worst case. The number 7 would omit the deepest level, and 8 counts only that level.

Question 29Fundamentals of Artificial Intelligence

Which property must a heuristic satisfy for greedy best-first search to behave differently from uniform-cost search?

  1. It estimates remaining distance to the goal rather than accumulated cost so far Correct answer.
  2. It always returns a value of zero for every state
  3. It stores the full closed list of expanded states
  4. It guarantees that every step cost equals one

Explanation

Greedy best-first search orders the frontier by an estimate of the cost still to come, so it ignores how expensive the path so far has been. Uniform-cost search instead orders by accumulated path cost g. A heuristic that estimates remaining distance is what makes the greedy ordering distinct.

Question 30Fundamentals of Artificial Intelligence

In constraint satisfaction, which technique removes values from a variable's domain that cannot participate in any consistent assignment with a single neighbouring variable?

  1. Forward checking
  2. Backtracking with conflict-directed backjumping
  3. Arc consistency Correct answer.
  4. Minimum remaining values ordering

Explanation

Arc consistency repeatedly deletes values from a variable's domain when no value in a neighbouring domain is compatible with them, until every arc is consistent. Forward checking is weaker because it only prunes after each assignment rather than propagating to a fixpoint. Variable ordering heuristics change search order but do not by themselves remove inconsistent values.

Question 31Fundamentals of Artificial Intelligence

A knowledge base uses default logic with the default 'birds typically fly'. Tweety is a bird and also a penguin, and penguins are stated not to fly. What does a well-designed default theory conclude?

  1. The knowledge base is inconsistent and entails everything
  2. Neither conclusion can be drawn because defaults are probabilistic
  3. Tweety flies, because the bird default applies first
  4. Tweety does not fly, because the more specific penguin rule blocks the default Correct answer.

Explanation

Default reasoning is non-monotonic: a conclusion holds only while nothing more specific contradicts it. The penguin rule is more specific and explicitly denies flight, so it defeats the general bird default. This specificity-based defeat is how such systems avoid drawing the conflicting conclusion.

Question 32Fundamentals of Artificial Intelligence

What does the naive Bayes classifier assume about the features given the class label?

  1. That the features are mutually independent conditional on the class Correct answer.
  2. That the features are always normally distributed within each class
  3. That the class labels are independent of one another for each example
  4. That every feature has the same variance across all classes

Explanation

Naive Bayes factorises the class-conditional joint probability into a product of per-feature probabilities, which is valid only under conditional independence given the class. The assumption is often violated in practice, yet the classifier frequently performs well. Distributional and variance assumptions are optional modelling choices, not the defining assumption.

Question 33Fundamentals of Artificial Intelligence

A decision tree is grown until every leaf is pure on the training set. Training accuracy is 100 percent while held-out accuracy is much lower. Which remedy most directly addresses this?

  1. Adding more features to the training data
  2. Removing the validation set from the pipeline
  3. Increasing the maximum depth of the tree
  4. Pruning the tree or limiting its depth Correct answer.

Explanation

A fully grown tree can memorise idiosyncrasies of the training sample, which is the classic overfitting pattern of high training accuracy and poor generalisation. Pruning or capping depth constrains model complexity so the tree captures broader regularities. Adding features or depth increases capacity and would worsen the gap.

Question 34Fundamentals of Artificial Intelligence

In reinforcement learning, what distinguishes a model-free method from a model-based method?

  1. Model-based methods only work in episodic tasks
  2. Model-free methods require a labelled dataset of optimal actions
  3. Model-based methods cannot use temporal-difference updates
  4. Model-free methods learn values or policies without explicitly estimating environment dynamics Correct answer.

Explanation

A model-based agent estimates transition and reward dynamics and can use them for planning, whereas a model-free agent learns value estimates or a policy directly from interaction. Both families can use temporal-difference learning and can operate in episodic or continuing tasks. Labelled optimal actions are a supervised-learning concept, not a requirement of model-free reinforcement learning.

Question 35Fundamentals of Artificial Intelligence

Which statement correctly contrasts supervised and unsupervised learning?

  1. Supervised learning always requires more computation than unsupervised learning
  2. Supervised learning uses labelled examples, while unsupervised learning seeks structure in unlabelled data Correct answer.
  3. Unsupervised learning cannot be evaluated by any quantitative measure
  4. Supervised learning never uses more than one input feature

Explanation

The defining difference is the presence of target labels: supervised methods learn a mapping from inputs to known outputs, while unsupervised methods such as clustering and dimensionality reduction look for structure without labels. Computational cost and feature count vary widely in both settings. Unsupervised results can be assessed with measures such as silhouette scores or reconstruction error.

Question 36Fundamentals of Artificial Intelligence

A perceptron is trained on data that is not linearly separable. What is the guaranteed outcome of the standard perceptron learning rule?

  1. It converges to the maximum-margin separator
  2. It may oscillate indefinitely without converging Correct answer.
  3. It converges to a separator with zero training error
  4. It always converges after exactly one pass over the data

Explanation

The perceptron convergence theorem guarantees termination only when the classes are linearly separable. Without separability the update rule can keep cycling, and the weights may never settle. Maximum-margin solutions come from support vector machines rather than the perceptron rule.

Question 37Fundamentals of Artificial Intelligence

Which sentence is a well-formed atomic formula in first-order logic?

  1. Likes(Alice, Bob) and Runs(Bob)
  2. Likes(Alice, Bob) implies Likes(Bob, Alice)
  3. Likes(Alice, Bob) Correct answer.
  4. forall x Likes(x)

Explanation

An atomic formula applies a predicate symbol to a tuple of terms, and Likes(Alice, Bob) has exactly that form with two constant arguments. The other candidates combine formulas with connectives or add a quantifier, which produces compound or quantified formulas rather than atoms.

Question 38Fundamentals of Artificial Intelligence

A rule-based system has the rules 'if fever then infection' and 'if infection then refer to clinic'. The fact 'fever' is asserted. Which inference does forward chaining produce first?

  1. Infection, because its premise matches a known fact Correct answer.
  2. Both conclusions simultaneously, because chaining is parallel
  3. Neither conclusion, because the rules form a cycle
  4. Refer to clinic, because it is the final conclusion

Explanation

Forward chaining is data driven: it scans rules whose premises are satisfied by current facts and fires them, adding their conclusions. With fever known, the first rule fires and adds infection, which then enables the second rule on the next cycle. Conclusions are therefore derived in successive steps rather than all at once.

Question 39Fundamentals of Artificial Intelligence

In a Markov decision process, what does the Markov property state about the transition to the next state?

  1. It depends on the current state and on the step index
  2. It is independent of the action selected in the current state
  3. It depends only on the current state and the chosen action Correct answer.
  4. It depends on the entire history of states and actions

Explanation

The Markov property says the future is conditionally independent of the past given the present state and action, so the transition distribution needs only those two arguments. This is what makes dynamic programming and many reinforcement-learning methods tractable. Dependence on full history or on a step index would violate the property.

Question 40Fundamentals of Artificial Intelligence

A planning domain has actions with preconditions and effects. An action's precondition is true in the current state, but one of its delete effects removes an atom that a later action needs. What does a correct forward state-space planner do?

  1. Apply the action only if the deleted atom is restored immediately
  2. Reject the action because it removes a needed atom
  3. Ignore the delete effect because preconditions were satisfied
  4. Apply the action and record the resulting state without the deleted atom Correct answer.

Explanation

A forward planner applies any action whose preconditions hold and computes the successor state by adding add effects and removing delete effects. Whether a removed atom is needed later is discovered during subsequent search, not used to veto the action. This faithful successor generation is what keeps the search sound.

Question 41Machine Learning and Model Development

Which type of machine learning uses labeled training data to learn a mapping from inputs to outputs?

  1. Supervised learning Correct answer.
  2. Reinforcement learning
  3. Unsupervised learning
  4. Clustering

Explanation

Supervised learning trains on examples where each input is paired with a known target label, so the model learns to predict outputs for new inputs. The presence of ground-truth labels during training is the defining characteristic that separates it from the other paradigms.

Question 42Machine Learning and Model Development

In the bias-variance decomposition, which situation describes high bias?

  1. Training and validation error are both near zero
  2. The model fits training data closely but generalizes poorly
  3. The model is highly sensitive to small changes in training data
  4. The model underfits because it is too simple for the data Correct answer.

Explanation

High bias means the model's assumptions are too restrictive, so it cannot capture the underlying pattern and underfits both training and validation data. Variance, by contrast, concerns sensitivity to sampling noise.

Question 43Machine Learning and Model Development

A dataset has 90 negative and 10 positive examples. A model that always predicts negative achieves what accuracy?

  1. 50 percent
  2. 90 percent Correct answer.
  3. 100 percent
  4. 10 percent

Explanation

Accuracy is the fraction of predictions that are correct. Predicting negative for every example is correct on the 90 negatives and wrong on the 10 positives, giving 90/100, or 90 percent.

Question 44Machine Learning and Model Development

What is the primary purpose of splitting data into training and test sets?

  1. To speed up the training algorithm
  2. To balance the class distribution in the data
  3. To reduce the number of features used by the model
  4. To estimate how well the model generalizes to unseen data Correct answer.

Explanation

Holding out a test set simulates unseen data so performance estimates reflect generalization rather than memorization. Training-set performance alone is optimistically biased because the model has already seen those examples.

Question 45Machine Learning and Model Development

Which activation function outputs values strictly between zero and one?

  1. Linear
  2. Sigmoid Correct answer.
  3. ReLU
  4. Tanh

Explanation

The sigmoid function maps any real input to the interval between zero and one through the logistic formula. This bounded range makes it useful for producing probabilities.

Question 46Machine Learning and Model Development

What does the learning rate control in gradient descent?

  1. The number of features used per update
  2. The fraction of data held out for validation
  3. The number of layers in the network
  4. The size of the step taken along the negative gradient Correct answer.

Explanation

The learning rate scales the gradient update, determining how far parameters move each iteration. Too large a value can cause divergence, while too small a value slows convergence.

Question 47Machine Learning and Model Development

Which technique reduces overfitting by randomly disabling a fraction of neurons during each training step?

  1. Batch normalization
  2. Dropout Correct answer.
  3. Data augmentation
  4. Weight decay

Explanation

Dropout randomly zeroes activations during training, which prevents units from co-adapting and acts as a form of regularization. At inference time the full network is used with appropriately scaled weights.

Question 48Machine Learning and Model Development

In k-fold cross-validation with k equal to 5, how many times is a model trained on a distinct training subset?

  1. 10
  2. 5 Correct answer.
  3. 1
  4. 2

Explanation

Each of the five folds serves once as the validation set while the remaining four folds form the training set, so five separate models are trained. The results are then averaged.

Question 49Machine Learning and Model Development

A model shows low training error but much higher validation error. Which remedy is most appropriate?

  1. Add regularization such as weight decay or dropout Correct answer.
  2. Decrease the size of the training set
  3. Increase model capacity by adding layers
  4. Remove the validation set and train on all data

Explanation

The gap between low training error and high validation error indicates overfitting, i.e., high variance. Regularization constrains model complexity and typically narrows that gap.

Question 50Machine Learning and Model Development

Which metric is most appropriate for a highly imbalanced binary classification problem where missing positive cases is costly?

  1. Accuracy
  2. Recall of the positive class Correct answer.
  3. R-squared
  4. Mean squared error

Explanation

Recall measures the fraction of actual positives that the model correctly identifies, directly reflecting the cost of missed positive cases. Accuracy can look high on imbalanced data even when positives are ignored.