301. Neural networks are not a type of machine learning.

A. True
B. False
Answer: B

302. About the following code: SVRikemet=" rbf , C=1e3, gamma=0.01), it means that we use linear
kernel function.
A. True
B. False
Answer: A

303. The training error decreases as the model complexity increases.

A. True
B. False
Answer: B

304. The smart speaker is one of the typical application scenarios of speech processing.
A. True
B. False
Answer: A

305. At present, Al is still in the narrow Al phase.

A. True
B. False
Answer: A

306, Tensor Flow is now one of the most popular deep learning frameworks.
A. True
B. False
Answer: A

307. Tensor Flow is an end-to-end open-source platform for machine learning and deep learning.
A. True
B. False
Answer: A

308. A recurrent neural network can capture dynamic information in serialized data.
A. True
B. False
Answer: A

309. The softmax function is often used as the output layer of a multiclass classification task.
A. True
B. False
Answer: A

310. An artificial neural network (ANN) is an information processing system designed to imitate the
structure and functions of human brain.
A. True
B. False
Answer: A

311. Huawei's Ascend Al chip is one of the typical NPUs.

A. True
B. False
Answer: A

312. Mind Spore’s vision and value are to lower the thineshald for Al development in the industry,
release the computing power of Ascend chips, and facilitate inclusive Al
A. True
B. False
Answer: B

313. Mind Spore can use only Huawei-developed Ascend processors for training and inference.
A. True
B. False
Answer: A

14. Knowledge graph can be applied to social networking, gene sequencing, and risk control,
True
False
Answer: A

315. HUAWEI HCAl Engine can easily integrate multiple artificial intelligence (Al) capabilities with
apps.
A. True
B. False
Answer: A

316. Factory A produces 40% of an automotive company’s products, and factory B accounts for 60%
The detective rate of factory A is 1%, and that of factory Bis 2% It vehicle is defective, what is the
possibility that the vehicle is produced by factory A?
A. 0.2
B. 0.25
C. 0.3
D. 0.5
Answer: B

317. When we use this code: Grid Search CV (SVR (), param_grid=parameters, cv=3), cv=3
means()?
A. SVR iterative three times O
B. 3 classification
C. Do a 3 fold cross validation?
D. The data is 3d
Answer:

318. Which of the following Is the problem type for "predicting that Ihe store revenue in the next
quarter is 200,000-300,000 or 400,000-500,000 based on héstorical data with fabets”?
A. Classification problem
B. Regression preform
C. Rate problem
D. Chartering problem
Answer: D
319. According to the American Society of Automotive Enginwers (SAD), which of the folowing are the
lewels of autonomous driving clasutied by the degree of dependence on the system?
A. 10-14
B. 10-15
C. 11-15
D. 11-14
Answer: B

320. Which of the following is not a feature of Tensor Flow 2.0?

A. Multi-lingual
B. Multi-platform
C. Distributed
D. Multi-core CPU acceleration
Answer: A

321. Which of the following methods can be used to view tensor dimensions in Tensor Flow 2.0?
A. dimens
B. dtype
C. device
D. ndim
Answer: B

322. What is the purpose of adding the dropout layer when creating a network layer?
A. Extracting image features
B. Functioning as a classifier
C. Preventing over fitting
D. Increasing the non-linearity
Answer: C

323. Which of the following is not included in built-in metrics of keras.metrics in Tensor Flow 2.07?
A. mean
B. Accuracy
C. Recall
D. sum
Answer: D

324. What is the meaning of the epochs parameter in model fit(mnisttrain.image,

mnist.train.labels.epochs=5)?
A. The whole training set will be divided into five parts.
B. The whole training set will be divided into six parts
C. All training sets will be trained five times.
D. All test sets will be tested five times.
Answer: B

325. Which of the following statements about optimizers is correct?

A. The learning rate of the momentum optimizer does not need to be set manually.
B. The learning rate of the AdaGrad optimizer is not automatically updated.
C. The momentum optimizer is updated with the same learning rate for each iteration.
D. There is no relationship between the Adam optimizer and AdaGrad optimizer.
Answer: C

326. If each layer of a neural network uses the sigmoid activation function and the number of layers is
large, which of the following problems will occur?
A. Gradient explosion problem
B. Vartshing graclient problem
C. Under fitting problem
D. Over tatting problem
Answer: B

327. Which of the following statements about the cross-entropy loss function is correct?
A. The cross-entropy loss function is often used in classification problems
B. The cross-entropy loss function is often used in regression problems
C. The cross-entropy loss function can be used for both classification and regression problems
D. The cross-entropy loss function can only be used for binary classification problems
Answer: D

328. Regularization is often used in deep learning. What problem is resolved by regularization in deep
learning?
A. Data imbalance problem
B. Vanishing gradient problem
C. XOR problem
D. Over fitting problem
Answer: D

329. Deep leaning neural network training requires a large amount of matrix computation. Generally,
hardware is required tor a Compiler to be capable of parallel computing. Which of the following
hardware devices can provide the parallel Computing capability?
A. CPU
B. DIMM
C. GPU
D. Mainboard
Answer: A

330. Which of the following statements about a standard RNN model is correct?
A. Information attenuates after long-term memory transmission.
B. The many-to-many model structure does not exist.
C. The one-to-one model structure does not exist.
D. The time orientation is not considered during back propagation.
Answer: D

331. Dataset expansion can effectively affect over fitting occurrence. Which of the two Bowing
statements about dataset expansion its correct?
A. A larger dataset indicates a lower over fitting probability,
B. A larger dataset indicates a higher over fitting probability,
C. A smaller dataset indicates a lower over fitting probability.
D. An enlarged of shrunk dataset indicates a lower over fitting probability
Answer: D

332. Which of the following statements about a Generative Adversarial Networks(GAN) is incorrect?
A. The generative adversarial network includes a generative model and a discriminative model.
B. The generative adversarial network cannot be used to generate new images.
C. Some discriminators require labeled samples.
D. The input of the generative model can be random numbers
Answer: C

333. Which of the following is the output range of the Sigmoid function? O
A. (0,+ꚙ)
B. (0,1)
C. E1,1]
D. [-1,0]
Answer: B

334. Which of the following optimizers can automatically adjust the learning rate?
A. SGD algorithm
B. Mini-batch gradient descent algorithm
C. Momentum optimizer
D. AdaGrad optimizer
Answer: A

335. Which of the following is not included in the logical architecture of Ascend Al processors?
A. Control CPU
B. Al computing engine
C. DVPP
D. GPU
Answer: C

336. Which of the following is the main application of the Ascend 310 chip serving as the basis of the
A. Atlas 200 DK?
B. Model training
C. Model building
D. Model inference
Answer: A

337. Mind Spore has made many optimizations to improve computing performance. Which of the
following is not Mind Spore’s optimization?
A Pipeline parallelization
B. Dedicated memory
C. On-device execution
Answer: A

338. The core of an Al framework, that is, one of decisive factors of the programming paredigm is the
automatic differentiation technology used by the Al framework. Which of the following method is used
by Mind Spore?
A. Graph method
B. Operator overloading
C. Source code conversion
Answer: B

339. Which one doesn't belong to Huawei El Essential Platform?

A. Model Arts
B. Huawei HiLens
C. Graph Engine Service (GES)
D. Huawei HiAl
Answer: C

340. Which of the following HUAWEI HiAl modules is supported by Huawei chips?
A. HiAI Service
B. HiAl Foundation
C. HiAl Framework
D. HiAl Engine
Answer: B
341. Which of the following models can be used for classification?
A. Linear regression
B. SVM
C. Logistic regression
D. Decision tree
Answer: B

342. From sklearn. linear model, we can import () modules?

A. Ridge CV
B. Lasso CV
C. Linear Regression
D. SVM
Answer: C

343. In which of the following cases should machine learning be used to resolve actual problems?
A. Rules ate complex of cannot be described, such as facial recognition and speech recognition.
B. The rules of the task will be changed with time, such as part-ot-speech tagging, new words or
meanings will be generated at any time.
C. Data distribution changes over time, requiring constant readaptation of programs, such as
forecasting
the trend of commodity sales
D. Problems have low complexity and a small scale,
Answer: C

344, Which of the following are common types of dirty data?

A. Duplicate values
B. Missing values
C. Values in incorrect formats
D. Logically incorrect values
Answer:A

345. Which of the following can be the basis for node division in a decision tree?
A. Gini coefficient
B. ID3
C. Information entropy
D. CART
Answer: D

346. Which of the following problems can be resolved by using machine learning algorithms?
A. Classification problem
B. Regression problem
C. Reinforcement problem
D. Clustering problem
Answer: A

347. What are the application scenarios of computer vision?

A. Action analysis
B. Smart album
C. Image search
D. Voice navigation
Answer: C
348. Mf the total number of parameters at a layer of a fully-connected neural network is 330, the
numbers of neurons at the upper layer and the current layer may be (), respectively.
A. 32and 10
B. 33.and 10
C. 9 and 33
D. 10 and 33
Answer: A

349. Which of the following methods can be used to check whether a data type is tensor in
TensorFlow 2.0?
A. is tensor
B. is instance
C. device
D. dtype
Answer:A

350. If a model has a high accuracy on the training set but tow accuracy on the test set, what methods
can be used to improve the model?
A. Increasing the data volume of the training set
B. Adding a regularization term after the loss function
C. Dropout method
D. Early stopping
Answer:B

351. Which of the following methods can be used to resolve data imbalance problems in deep
learning tasks?
A. Random under sampling
B. Random oversampling
C. Synthetic sampling
D. Batch deletion
Answer:A

352. Deep teaming is 4 popular Al technology; Various problems may occur in deep learning tasks.
Which of the following problems may occur in deep learning tasks?
A Data imbalance problem
B. Overfitting problem
C. Vanishing gradient problem
D. Exploding gradient problem
Answer:c

353. Which of the following steps in deep learning are automatically performed by models?
A. Task analysis and locating
B. Model training
C. Feature extraction
D. Feature selection
Answer:D

354. Which of the following statements about a convolutional neural network pooling layer are
correct?
A. The pooling layer provides the function of dimension reduction.
B. Max-pooling and average-pooling are common pooling methods.
C. Featured images become smaller after pooling.
D. Pooling operations are implemented through window scanning.
Answer:D

355, Which of the following Atlas products are used in training scenarios?
A. Atlas 300 (model 9000)
B. Atlas 500
C. Atlas 800 (model 9000)
D. Atlas 900
Answer:A

356, Which of the following are common tensor operations in Mind Spore?
A. asnumpy ()
B. sized ()
C. switch ()
Answer:A

357. Which two of the following are included in Mind Spore?

A. TBE (operator development)
B. Mind Expression
C. Graph Engine
D. Third-party framework TF/Cafes
Answer:A,C,D

358. HUAWEI CLOUD Model Arts is 4 one-stop Al development platform. Which of the following
functions can be provided by Model Arts?
A. Data governance
B. Automatic learning
C. Visualized workflows
D. Al market
Answer:B

359. What are the major application scenarios of ModelArts?

A. Image
B. Speech
C. Text
D. Industrial control
Answer:A

360. What are challenges in Al capability development and application?

A. High barrier to entry
B. Low efficiency
C. Diverse requests
D. High computing power
Answer:A

361. The training error decreases as the model complexity increases.

A. True
B. False
Answer: B

362. The bottom layer of the random forest algorithm of an integration model is the ID3 tree.
A. True
B. False
Answer: A

363. In polynomial regression, a square term exists in a model's formula. Therefore, the model is
nonlinear.
A. True
B. False
Answer: A

364. At present, Al is still in the narrow Al phase.

A. True
B. False
Answer: A

365. Computer vision is a science of how to make computers “see".

A. True
B. False
Answer: A

366. TensorFlow is now one of the most popular deep learning frameworks
A. True
B. False
Answer: A

367. Porch & 4 Python-based machine Rarning computing framework released by Facebook. It is
developed bated on Torch, 4 scientific computing framework supported by a large number of machine
learning algorithms.
A. True
B. False
Answer: A

368. All convolutional kernels of the same convolutional layer on a convolutional neural network
share a weight.
A. True
B. False
Answer: A

369. An activation function is used to convert linear mapping into nonlinear mapping.
True
False
Answer: B

370. A recurrent neural network can capture dynamic information in serialized data.
A. True
B. False
Answer: A

371. The CPU improves Al performance by adding instructions and cores

A. True
B. False
Answer: A

372. The typical application development process of the application development service includes
mode! Evaluation.
A. True
B. False
Answer: A

373. Mind Spore supports interconnection with third-party chips and helps developers increase Mind
Spore application scenarios and expand the Al ecosystem.
A. True
B. False
Answer: A

374. Knowledge graph can be applied to social networking, gene sequencing, and risk control.
A. True
B. False
Answer: A

375. The HIAI mobile computing platform provides various upper-layer function service APIs to
ensure efficient running on mobile devices.
A. True
B. False
Answer: B

376, In machine learning, () needs to be input to a model to train the model and make predictions.
A. Artificial programs
B. Historical data
C. Neural networks
D. Training algorithms
ANSWER : B

377. For multiclass problems, softmax assigns probability values to each class. The sum of these
probability values is ( ).
A . 0.1
B. 0.5
C. 1
D . 100

ANSWER : C

378. What is the most important difference among batch gradient descent, mint-batch gradient
descent, and stochastic gradient descent?
A. Learning rate
B Gradient direction
C Batch size
D Gradient size

ANSWER : A
379. The three phases of Al include computing intelligence, ( ), and cognitive intelligence.
A. Behavior intelligence
B. Perceptual intelligence
C. Narrow Al
D. Strong Al

ANSWER : A,B

380, What is the purpose of adding the dropout layer when creating a network layer?
A. Extracting image features
B. Functioning as a classifier
C. Preventing overfitting
D. Increasing the non-linearity

ANSWER : C

381. What is the function of keras.preprocessing of TensorFlow 2.0?

A. Keras data processing tool
B. Keras model generation tool
C. Keras model deployment tool
D. Keras built-in optimizer

ANSWER : A

382. Which of the following methods can be used to view tensor dimensions in TensorFlow 2.0?
A. dimens
B. dtype
C. device
D. ndim

ANSWER : D

383. What is the meaning of the epochs parameter in model.fit(mnist train image,
mnist.train.labels.epochs=5)?
A. The whole training set will be divided into five parts.
B. The whole training set will be divided into six parts.
C. All training sets will be trained five times.
D. All test sets will be tested five times.

ANSWER : C

384. Which company first launched PyTorch?

A. Google
B. Huawei
C. Baidu
D. Facebook

ANSWER :D

385. Which of the following is the output range of the ReLU function?
A. [0,+2)
B. (0,1)
C [1,1]
D. [-1,0]

ANSWER : D

386 .Which of the following statements about backpropagation is incorrect?

A. Backpropagation means that errors are propagated in backward direction through a network.
B. Backpropagation is used to update the weights in neural network .
C. Backpropagation can only be used on a feedforward neural network.
D. Backpropagation can collaborate with a gradient descent algorithm to update a network weight
ANSWER : D
.
387. each layer of a neural network uses the Sigmoid activation function and the number of layers is
large, which of the following problems will occur?
A. Gradient explosion protein
B Vanishing gradient pattern
C Underfitting problem
D. Overfitting problem

ANSWER : B

388. Which of the following statements about the feedforward neural network is correct?
A. An Input node of the feedforward neural network has a computing capability.
B. Neurons at the same layer are interconnected.
C. Inter-layer information is transmitted in only one direction.
D. Neurons with computing capabilities connect the upper and lower layers

ANSWER :
.
389. Which of the following statements about a recurrent neural network is incorrect?
A. The short form of recurrent neural network is RNN.
B. A recurrent neural network can be expanded according to the time axis.
C. Long short-term memory (LSTM) is a recurrent neural network.
D. GRU is not a recurrent neural network.

ANSWER : C

390, In 1958, Rosentsatt invented the perceptron algorithm, which was used for about 10 years. In
1969, Minsky questioned the capability of the perceptron to resolve nonlinear problems and proposed
the farnous problern that slowed down the development of perceptron. The problem is an
A . AND problem
B. OR problem
C. XOR problem
D. XAND problem

ANSWER : D