Huawei Certified ICT Associate - AI — Free Practice Questions
10 free sample questions from a bank of 89, with the correct answers and explanations. No signup required — start practising right now.
1Which of the following options is not central to linear algebra?
Matrix theory
Vector space
Linear transformation
Probability theory
Answer: D
The short version
D — Probability theory is the outsider. Linear algebra is built on matrices, vector spaces, and linear maps, while probability theory is a separate branch of mathematics.
Key concepts in this question
Linear algebra core: matrices, vector spaces, and linear transformations form the central objects of the subject.
Probability theory: the study of randomness and distributions, which uses linear algebra but is not part of it.
HCIA-AI math basis: the exam separates linear algebra, probability/statistics, and optimization as distinct foundations.
Why D is correct
Matrix theory gives the computational machinery of linear algebra, vector spaces give its abstract setting, and linear transformations are the structure-preserving maps it studies. Probability theory — random variables, distributions, expectation — belongs to probability and statistics, not to linear algebra itself, so it is the option that is not central.
Why the others are wrong
A. Matrix theory is a pillar of linear algebra: systems of equations, transformations, and eigenvalues are all expressed with matrices.
B. Vector spaces are the fundamental objects of linear algebra; the whole subject generalizes their structure.
C. Linear transformations are the morphisms of linear algebra and the reason matrices and vector spaces matter.
H13-311 exam tip
If three options name matrices, spaces, and maps, the probability-flavored option is the designed distractor.
2According to the definition of information entropy, what is the bit entropy of throwing a uniform coin?
0.5
-1
1
0
Answer: C
The short version
C — A fair coin carries exactly 1 bit of entropy. With two equally likely outcomes, Shannon entropy is one binary unit of uncertainty.
Key concepts in this question
Shannon entropy: H = −Σ p(x) log₂ p(x), measured in bits when the log is base 2.
Uniform binary outcome: p(heads) = p(tails) = 0.5 maximizes uncertainty for a one-bit event.
Bit as unit: 1 bit is the information gained by resolving one fair yes/no choice.
Why C is correct
For a fair coin, H = −(0.5 × log₂ 0.5 + 0.5 × log₂ 0.5) = −(0.5 × −1 + 0.5 × −1) = 1. Each toss resolves exactly one binary unit of uncertainty, so the bit entropy is 1, matching the intuitive idea that one coin flip stores or conveys one bit.
Why the others are wrong
A. 0.5 would be the probability of each face, not the entropy; entropy sums over both outcomes and equals 1.
B. −1 is the value of log₂ 0.5, a single term inside the sum — entropy negates and sums these, giving a positive 1.
D. 0 describes a certain event with no uncertainty, the opposite of a fair coin toss.
H13-311 exam tip
Memorize the anchor: fair coin = 1 bit; certain outcome = 0 bits; then compute from H = −Σ p log₂ p.
3In a neural network, knowing the weight and deviations of each neuron is the most important step. If you know the exact weights and deviations of neurons in some way, you can approximate any function.What is the best way to achieve this?
Random assignment, pray that they are correct
The above is not correct
Search for a combination of weight and deviation until the best value is obtained
Assign an initial value to iteratively update weight by checking the difference between the best value and the initial.
Answer: D
The short version
D — Train iteratively from an initialization. Weights are found by starting somewhere reasonable and repeatedly updating against the error, not by guessing once or exhaustively searching.
Key concepts in this question
Weight initialization: starting values break symmetry before training begins.
Loss and gradients: the gap between prediction and target guides each update via backpropagation.
Iterative optimization: gradient-descent-style steps gradually move weights toward values that approximate the target function.
Why D is correct
Neural networks are universal approximators only if their weights can actually be found. Random guessing or brute-force search over a huge continuous weight space is infeasible, so the practical method is to assign initial values and iteratively update them by measuring the difference between current output and the desired output and stepping the weights to reduce it.
Why the others are wrong
A. Pure random assignment with no update step leaves accuracy to chance and is not a training method.
B. Dismissing all approaches answers nothing; iterative optimization is the established correct approach.
C. Exhaustively searching every weight-bias combination is computationally impossible for real networks, which is why gradient-based iteration is used instead.
H13-311 exam tip
Any option describing init plus error-driven iterative update is the training answer; pure guessing or brute force is always wrong.
4What are the service categories included in Huawei Cloud EI Enterprise Intelligence? (Multiple Choice)
EI speech semantics
EI visual cognition
EI online games
EI industry scene
Answer: A, B, D
The short version
A, B and D — Speech/semantics, visual cognition, and industry scenarios are EI service categories. Online games are not an EI Enterprise Intelligence category.
Key concepts in this question
EI speech semantics: speech recognition, synthesis, and natural language understanding services.
EI visual cognition: image and video recognition and analysis services.
EI industry scenarios: packaged solutions for verticals such as manufacturing, logistics, and cities.
Why A, B and D are correct
Huawei Cloud EI (Enterprise Intelligence) product documentation groups its AI services into speech/semantics, visual cognition (computer vision), and industry-scenario solutions that apply those capabilities to real business domains. These three are the documented EI service families, so the banked combination A, B and D is the approved multi-select answer.
Why the others are wrong
C. Online games are an entertainment application vertical, not an EI service category; nothing in the EI portfolio is sold as EI online games.
H13-311 exam tip
Remember EI as see plus speak plus solve: vision, speech/semantics, and industry scenarios — anything gaming-flavored is the distractor.
5Which of the following is not an artificial intelligence school?
Statisticalism
Behaviorism
Symbolism
Connectionism
Answer: A
The short version
A — Statisticalism is not one of the classic AI schools. The textbook trio is symbolism, connectionism, and behaviorism.
Key concepts in this question
Symbolism: intelligence through symbols and logical rules, the classic knowledge-based school.
Connectionism: intelligence emerging from interconnected neuron-like units, the neural network school.
Behaviorism: intelligence judged and built through interaction and behavior with the environment.
Why A is correct
HCIA-AI teaching follows the classic division of AI into three schools: symbolism (logic and knowledge), connectionism (neural networks), and behaviorism (agents acting and adapting). Statistical methods are tools used across AI, but Statisticalism is not counted as one of these three foundational schools, making A the option that does not belong.
Why the others are wrong
B. Behaviorism is one of the three recognized schools, focused on perception-action and adaptive behavior.
C. Symbolism is the founding school of logic-based and expert-system AI.
D. Connectionism is the school behind neural networks and modern deep learning.
H13-311 exam tip
Recite the trio as symbols, neurons, behaviors — any fourth -ism on the exam is the planted outsider.
6Which of the following options is not a reason for traditional machine learning algorithms to promote the development of deep learning?
Dimensional disaster
Local invariance and smooth regularization
Manifold learning
Feature Engineering
Answer: D
The short version
D — Feature engineering is the manual practice deep learning replaces, not a driving theory. The cited theoretical drivers are dimensionality, invariance/regularization, and manifold structure.
Key concepts in this question
Curse of dimensionality: classical methods degrade as input dimensions grow.
Local invariance and smooth regularization: assumptions that break down on complex real data.
Manifold learning: the idea that high-dimensional data concentrates near lower-dimensional structure deep models can capture.
Why D is correct
The curriculum presents the push toward deep learning as motivated by theoretical limits of shallow methods: the dimensional disaster, the limits of local-invariance and smoothness priors, and manifold structure in real data. Feature engineering is the labor-intensive manual workflow of traditional machine learning — the symptom being escaped — rather than one of these theoretical reasons, so it is the banked not-a-reason answer.
Why the others are wrong
A. The curse of dimensionality is a core cited limitation motivating distributed deep representations.
B. Local invariance and smooth regularization assumptions are presented as shallow-method priors that fail on complex tasks.
C. Manifold learning arguments are used to justify why deep architectures can model real-world data geometry.
H13-311 exam tip
Separate theories from workflow: dimensionality, invariance, manifold are the theory distractors that count as reasons; the engineering-labor option is the answer.
7Which of the following options is not the session mode used by TensorFlow?
Explicitly call the session to generate function
Explicitly call the session to close function
Multiple POST queries
Through the Python context manager
Answer: C
The short version
C — Multiple POST queries is not a TensorFlow session mode. Sessions are managed by explicit calls or a Python context manager.
Key concepts in this question
TensorFlow 1.x session: the runtime object that executes a computation graph.
Explicit session management: creating a session and closing it with dedicated calls.
Context manager: the with tf.Session() pattern that closes the session automatically.
Why C is correct
In classic TensorFlow, a session is generated with an explicit creation call, closed with an explicit close call, or scoped through a Python context manager (with-block) that handles cleanup. Issuing multiple POST queries is a web-API notion, not a session lifecycle mode, so it is the option that does not belong.
Why the others are wrong
A. Explicitly creating (generating) a session is the standard documented first step of session use.
B. Explicitly closing a session to release resources is the standard documented cleanup step.
D. Managing the session through a Python context manager is the idiomatic pattern taught in TensorFlow materials.
H13-311 exam tip
Session answers always involve create, close, or with-block — any web-request-flavored option is the distractor.
8Which of the following are solutions for the Huawei Cloud EI industry scenario? (Multiple Choice)
Intelligent Logistics
Intelligent Finance
Intelligent transportation
Intelligent Water
Intelligent manufacturing
Answer: A, B, C, D, E
The short version
A, B, C, D and E — All five listed verticals are EI industry-scenario solutions. Logistics, finance, transportation, water, and manufacturing are all packaged EI offerings.
Key concepts in this question
EI industry scenarios: pre-packaged AI solutions that apply EI capabilities to a specific vertical.
Vertical coverage: Huawei documents EI solutions across transport, finance, industry, utilities, and logistics.
All-of-the-above pattern: when every listed vertical matches the portfolio, the full combination is correct.
Why A, B, C, D and E are correct
Huawei Cloud EI industry scenarios include intelligent logistics, intelligent finance, intelligent transportation, intelligent water management, and intelligent manufacturing. Since each option names a genuine documented EI vertical solution, the banked answer selects all five, and no option needs to be excluded.
Why the others are wrong
A. Intelligent logistics is a documented EI scenario, so it must be included, not excluded.
B. Intelligent finance is a documented EI scenario, so it must be included, not excluded.
C. Intelligent transportation is a documented EI scenario, so it must be included, not excluded.
D. Intelligent water is a documented EI scenario, so it must be included, not excluded.
E. Intelligent manufacturing is a documented EI scenario, so it must be included, not excluded.
H13-311 exam tip
For EI scenario lists, default to including every intelligent-plus-industry option unless one is obviously non-industrial.
9Python regular expressions are a special sequence of characters that makes it easy to check if a string matches a pattern.
True
False
Answer: A
The short version
A — True: that is exactly what Python regular expressions are. They are special character sequences for testing whether strings match a pattern.
Key concepts in this question
Regular expression: a compact pattern language for describing sets of strings.
Pattern matching: checking, searching, and extracting substrings that fit the pattern.
Python re module: the standard library that implements regex matching and substitution.
Why A is correct
The stem gives the textbook definition of a regular expression: a special sequence of characters that makes it easy to check whether a string matches a pattern. Python implements this through its re module for match, search, split, and substitution, so the statement is true.
Why the others are wrong
B. False would deny the standard definition; regexes exist precisely to test strings against patterns, so False contradicts both the definition and the re module.
H13-311 exam tip
Definition-style true/false stems that match the textbook sentence word-for-word are almost always True.
10What of the following does belong to convolutional neural network (CNN)? (Multiple Choice)
AlexNet
VGGNet
GoogleNet
ResNet
Answer: A, B, C, D
The short version
A, B, C and D — All four are convolutional neural networks. AlexNet, VGGNet, GoogLeNet, and ResNet are landmark CNN image models.
Key concepts in this question
Convolutional neural network: a network built around convolutional layers that learn spatial feature hierarchies.
ImageNet-era landmarks: each listed model advanced CNN design — depth, small filters, inception modules, residual links.
Shared family: despite different blocks, all four process images with stacked convolution and pooling.
Why A, B, C and D are correct
AlexNet reignited CNN research on ImageNet, VGGNet showed the power of deep stacks of small filters, GoogLeNet introduced inception modules within a convolutional architecture, and ResNet added residual connections to train very deep convolutional nets. All four are canonical CNNs, so the banked answer includes every option.
Why the others are wrong
A. AlexNet is the landmark CNN breakthrough, so it must be included, not excluded.
B. VGGNet is a defining deep CNN family, so it must be included, not excluded.
C. GoogLeNet is an inception-based CNN, so it must be included, not excluded.
D. ResNet is a residual CNN and one of the most cited vision backbones, so it must be included, not excluded.
H13-311 exam tip
AlexNet, VGG, GoogLeNet, ResNet form the must-know CNN quartet — if all four appear, take all four.