Difficulty: Hard
Under the Markov property, what determines the next state?
A Reward totals only
B Full history only
C Current state and action
D Future states only
Answer: C. Current state and action. In an MDP, the next-state distribution depends on the current state and action.
What does an optimal policy map each state to?
A An expected-return maximizing action
B A test partition
C A training label
D A nearest centroid
Answer: A. An expected-return maximizing action. A policy maps states to actions that maximize expected return.
Bellman equations relate immediate reward to which quantity?
A Training-set accuracy
B Search depth only
C Feature count
D Discounted future value
Answer: D. Discounted future value. Bellman equations express value recursively through reward and future value.
The discount factor gamma controls what?
A Size of the state space
B Weight of future rewards
C Noise in observations
D Number of training labels
Answer: B. Weight of future rewards. Discounting determines how strongly future rewards contribute to return.
Which pair does a Q-value estimate return for?
A Feature and label
B State and action
C Model and dataset
D Search and goal
Answer: B. State and action. Q-values estimate expected return after taking an action in a state.
During Q-learning, the bootstrap target uses which next-state value?
A First action only
B Average training loss
C Minimum feature distance
D Maximum action value
Answer: D. Maximum action value. Q-learning uses the maximum estimated next-action value in its target.
SARSA updates from which next action?
A The maximum possible action
B The action actually selected
C The previous episode’s action
D A random test action
Answer: B. The action actually selected. SARSA is on-policy and uses the next action selected by its current policy.
Besides hidden states, what does an HMM model?
A Class centroids
B Tree branches
C Reward labels
D Observation emissions
Answer: D. Observation emissions. An HMM models transitions among hidden states and emissions of observations.
The Viterbi algorithm finds which sequence?
A Shortest feature list
B Best cluster centers
C Largest reward total
D Most likely hidden states
Answer: D. Most likely hidden states. Viterbi decoding finds the most likely state sequence for observations.
For an HMM, what does the forward algorithm compute?
A Observation-sequence probability
B Most likely single state
C A decision-tree split
D A cluster count
Answer: A. Observation-sequence probability. The forward algorithm sums path probabilities to compute sequence likelihood.
Transformer self-attention lets tokens do what?
A Choose class labels
B Set training rewards
C Weight other token relevance
D Store full image pixels
Answer: C. Weight other token relevance. Self-attention assigns relevance weights between tokens in a sequence.
Positional encodings provide information about what?
A Reward magnitude
B Image brightness
C Class balance
D Token order
Answer: D. Token order. Positional encodings let a transformer represent sequence positions.
Autoregressive language models commonly predict what?
A A pixel boundary
B The next token
C A reward function
D A cluster centroid
Answer: B. The next token. Autoregressive language modeling predicts upcoming tokens from preceding context.
How does an embedding represent a token?
A A raw sentence
B A numeric vector
C A decision tree
D A reward table
Answer: B. A numeric vector. Embeddings map discrete tokens to learned numeric vectors.
Convolutional filters are useful for detecting what?
A Search goals
B Class frequencies only
C Local input patterns
D Future rewards
Answer: C. Local input patterns. Convolutional kernels respond to local patterns such as edges in image regions.
Object detection commonly outputs what?
A Classes and bounding boxes
B A reward table
C A token sequence
D Only one image label
Answer: A. Classes and bounding boxes. Object detection identifies object classes and their locations in an image.
Semantic segmentation assigns which output?
A A token to each word
B A box to each object
C A class to each pixel
D One class to each image
Answer: C. A class to each pixel. Semantic segmentation labels pixels according to their semantic class.
Intersection over Union compares which regions?
A Predicted and true regions
B Classes and labels
C Training and test sets
D States and actions
Answer: A. Predicted and true regions. IoU divides the overlap area by the union area of predicted and true regions.
In a generative adversarial network, which model creates samples?
A Generator
B Discriminator
C Tokenizer
D Optimizer
Answer: A. Generator. The generator produces candidate samples for the discriminator to assess.
Model cards should document which details?
A Only user passwords
B Only training duration
C Only source-code length
D Uses, evaluation, limitations
Answer: D. Uses, evaluation, limitations. Model cards communicate intended uses and evaluation details, including limitations.
Which four functions form NIST AI RMF Core?
A Collect, Label, Store, Share
B Plan, Build, Run, Review
C Govern, Map, Measure, Manage
D Train, Test, Ship, Monitor
Answer: C. Govern, Map, Measure, Manage. NIST AI RMF Core contains GOVERN, MAP, MEASURE, and MANAGE.
NIST identifies which trait as part of trustworthy AI?
A Guaranteed unbiased outputs
B Zero-resource operation
C Valid and reliable
D Always fully autonomous
Answer: C. Valid and reliable. NIST lists validity and reliability among trustworthy-AI characteristics.
Retrieval-augmented generation adds what before text generation?
A A random search path
B Retrieved external context
C A larger output image
D A new classifier label
Answer: B. Retrieved external context. RAG retrieves relevant external information and supplies it as context to generation.
Distribution shift means what changed?
A Deployment data distribution
B Training algorithm name
C Number of output classes
D Model file format
Answer: A. Deployment data distribution. Performance may fall when deployment data differs from training data.
By NIST terminology, what is an adversarial example?
A A model’s reward table
B A harmless feature scale
C A random training label
D Input crafted to mislead
Answer: D. Input crafted to mislead. NIST defines an adversarial example as a modified sample that induces model misclassification or misbehavior.