Huawei HCIA-AI H13-311 Practice Exam

• A. Evolutionism
• B. Behaviorism
• C. Symbolism
• D. Connectionism Correct answer

Option D is correct because connectionism models intelligence as networks of simple interconnected units, which is exactly what artificial neural networks do. Symbolism relies on explicit logic and rules, evolutionism on evolutionary search, and behaviorism on learning through interaction with the environment.

• A. True
• B. False Correct answer

False. DeepSeek-V3 is a Mixture-of-Experts (MoE) model: it has a very large total parameter count, but a router activates only a small subset of experts, and so only a fraction of the parameters, for each token. A dense model is the one that activates all parameters for every token.

• A. Machine translation
• B. Speech synthesis
• C. Semantic segmentation of images Correct answer
• D. Image classification Correct answer

Options C and D are correct because semantic segmentation and image classification both analyze images, which is the core of computer vision. Machine translation (A) is a natural language processing task and speech synthesis (B) belongs to speech processing.

• A. Herbert Simon
• B. John McCarthy
• C. Alan Turing Correct answer
• D. Marvin Minsky

Option C, Alan Turing, is correct because he proposed the Turing test in his 1950 paper to judge whether a machine can show human-like intelligence in conversation. McCarthy (B) coined the term artificial intelligence, and Simon (A) and Minsky (D) were other early AI pioneers who did not propose this test.

• A. Filtering spam using labeled emails
• B. Recognizing digits from labeled images
• C. Predicting house prices from past sales
• D. Grouping unlabeled documents by topic Correct answer

Option D is correct because grouping unlabeled documents by topic is clustering, which finds structure in data with no labels. Spam filtering and digit recognition are supervised classification on labeled data, and predicting house prices from past sales is supervised regression.

• A. Validation set Correct answer
• B. Feature set
• C. Test set
• D. Training set

Option A is correct because the validation set is held out from training and used to compare hyperparameter settings and choose a model. The test set is kept for the final, unbiased evaluation, the training set is used to fit parameters, and a feature set is just the input variables, not a data split.

• A. Each new learner in boosting focuses on earlier prediction errors Correct answer
• B. Random forest combines decision trees trained with bagging Correct answer
• C. Bagging trains base learners independently on bootstrap samples Correct answer
• D. Boosting trains all base learners in parallel on bootstrap samples
• E. Random forest builds its trees one after another from residuals

Options A, B and C are correct. Boosting trains learners sequentially, each concentrating on the errors of the earlier ones, and bagging trains base learners independently on bootstrap samples; random forest is bagged decision trees with random feature selection. Option D is wrong because boosting is sequential, not parallel, and option E is wrong because building trees one after another from residuals describes gradient boosting, not random forest.

• A. True Correct answer
• B. False

True. A model that cannot fit even the training data shows high error on the training set, and it also performs badly on validation data. That pattern signals underfitting (high bias). Overfitting shows a low training error with a high validation error.

• A. ReLU
• B. Tanh Correct answer
• C. Sigmoid
• D. Softmax

Option B is correct because tanh squashes its input into the range (-1, 1). ReLU outputs values from 0 to infinity, sigmoid outputs values in (0, 1), and softmax outputs values in (0, 1) that sum to 1 across classes.

• A. 28×28
• B. 12×12
• C. 16×16
• D. 14×14 Correct answer

Option D is correct. With no padding and stride 1, the 5×5 convolution gives (32 - 5) + 1 = 28, so the map is 28×28. The 2×2 max pooling with stride 2 then halves it to 14×14. Option A ignores the pooling step, and options B and C do not follow from either calculation.

• A. Training must process tokens one at a time, as in an RNN
• B. Multi-head attention shares one set of weights across all heads
• C. Self-attention processes all positions of a sequence in parallel Correct answer
• D. Positional encoding supplies token order information Correct answer

Options C and D are correct. Self-attention relates all positions of a sequence at once, so training is parallel, and because attention itself ignores order, positional encoding is added to supply token order. Option A is wrong because token-by-token processing is how an RNN works, and option B is wrong because each attention head has its own projection weights.

• A. Neuron counts for each layer
• B. Batch sizes that fit memory
• C. Gradients of the loss function Correct answer
• D. Optimal learning rates per epoch

Option C is correct because automatic differentiation computes the gradients of the loss with respect to the parameters, which backpropagation then uses to update them. It does not choose layer sizes, batch sizes or learning rates, which are set by the developer or by hyperparameter search.

• A. 0
• B. 4
• C. 8 Correct answer
• D. 16

Option C is correct. For y = x², the gradient is dy/dx = 2x = 4 at x = 2. PyTorch accumulates gradients into x.grad on each backward() call unless they are zeroed, so after two calls x.grad = 4 + 4 = 8. Option B would be the value after a single call.

• A. Converting the trained model to an inference format
• B. Extracting and normalizing features from raw records Correct answer
• C. Labeling samples with target categories Correct answer
• D. Searching for the best learning rate
• E. Removing duplicate and corrupted samples Correct answer

Options B, C and E are correct because extracting and normalizing features, labeling samples and cleaning duplicate or corrupted data all prepare the dataset before training. Converting a trained model to an inference format belongs to deployment, and searching for the best learning rate is hyperparameter tuning during model training.

• A. LoRA Correct answer
• B. Model distillation
• C. Full-parameter fine-tuning
• D. Prefix tuning

Option A is correct because LoRA (Low-Rank Adaptation) freezes the pre-trained weights and trains small low-rank matrices added to them. Full-parameter fine-tuning updates all weights, distillation trains a smaller student model from a larger teacher, and prefix tuning trains prefix vectors rather than low-rank matrices.

• A. Adding worked examples to the prompt Correct answer
• B. Asking the model to reason step by step Correct answer
• C. Training low-rank adapters on domain data
• D. Continuing pre-training on a domain corpus
• E. Assigning the model a role in the system prompt Correct answer

Options A, B and E are correct because few-shot examples, step-by-step reasoning prompts and a system-prompt role all steer the output through the input alone, leaving the weights unchanged. Training low-rank adapters and continuing pre-training on a domain corpus both update weights.

• A. To cap the number of tokens in each model reply
• B. To store earlier user questions for later retrieval
• C. To set the model's role, tone and rules for the session Correct answer
• D. To update the model weights using the developer's examples

Option C is correct because a system prompt is the instruction block that sets the model's role, tone and rules for the whole session. Option A describes a generation parameter such as the maximum token setting, option B describes conversation memory or retrieval, and option D describes fine-tuning, which a prompt never does because the weights stay unchanged.

• A. True Correct answer
• B. False

True. User-based collaborative filtering finds users whose past behavior, such as ratings or clicks, is similar to the target user's, then recommends the items those similar users liked that the target user has not yet seen.

• A. Dialogue management
• B. Natural language understanding
• C. Speech synthesis
• D. Automatic speech recognition Correct answer

Option D is correct because automatic speech recognition (ASR) is the stage that turns the audio signal into text. Natural language understanding and dialogue management work on that text afterwards, and speech synthesis goes the other way, turning reply text into audio.

• A. Rank candidate transcripts by acoustic score
• B. Detect the wake word in streaming audio
• C. Convert the reply text into spoken audio
• D. Extract parameters like time and party size Correct answer

Option D is correct because slot filling pulls the parameters a task needs, such as time or party size, out of the user's utterance so the assistant can act on the intent. Ranking transcripts is an ASR decoding task, wake-word detection is a separate front-end task, and converting text to audio is speech synthesis.