Wednesday , November 18 2026
Breaking News

Deep Learning Certification Exam Answers 2026 (All 100 Questions)

Deep Learning Certification exam answers 2026: This page lists all 100 questions of the official Deep Learning Certification exam by Itronix Solutions, each with its four answer options and the correct answer highlighted in green. Use it to revise every topic before you attempt the test and earn your verifiable certificate.

About this certification: Deep learning is a type of machine learning and artificial intelligence (AI) that imitates the way humans gain certain types of knowledge.

Certification Deep Learning Certification
Provider Itronix Solutions
Category AI & Machine Learning
Total questions 100 multiple-choice questions
Marks 1 mark per question (100 marks)
Passing score 80% (80 out of 100)
Retake Available if you do not pass
Certificate Verifiable online certificate

Take the exam: Deep Learning Certification on Itronix Solutions. See every other Itronix exam in our complete list of Itronix certification answers.

Deep Learning Exam Questions and Answers

Questions 1 to 25

1. When did deep learning start?

A. 1943 ✔
B. 1978
C. 1962
D. 1989

Correct answer: A. 1943

2. Who is father of deep learning?

A. David Rumelhart
B. Frank Rosenblatt ✔
C. Alex Krizhevsky
D. Ilya Sutskever

Correct answer: B. Frank Rosenblatt

3. Which neural network has only one hidden layer between the input and output?

A. Deep neural network
B. Feed-forward neural networks
C. Shallow neural network ✔
D. Recurrent neural networks

Correct answer: C. Shallow neural network

4. Deep learning algorithms are _______ more accurate than machine learning algorithms in image classification

A. 33%
B. 0.4%
C. 0.41% ✔
D. 0.37%

Correct answer: C. 0.41%

5. How many layers of Deep learning algorithms are constructed?

A. 1
B. 2
C. 3
D. 4 ✔

Correct answer: D. 4

6. The Deep Learning first layer is called the?

A. inner layer ✔
B. hidden layer
C. outer layer
D. None of the above

Correct answer: A. inner layer

7. Assume that your machine has a large enough RAM dedicated to training neural networks. Compared to using stochastic gradient descent for your optimization, choosing a batch size that fits your RAM will lead to

A. a less precise and slower update.
B. a more precise but slower update. ✔
C. a less precise but faster update.
D. a more precise and faster update.

Correct answer: B. a more precise but slower update.

8. Which of the following methods DOES NOT prevent a model from overfitting to the training set?

A. Pooling ✔
B. Early stopping
C. Data augmentation
D. Dropout

Correct answer: A. Pooling

9. Which of the following is/are Limitations of deep learning?

A. Data labeling
B. Obtain huge training datasets
C. None
D. All of above ✔

Correct answer: D. All of above

10. Which of the following is well suited for perceptual tasks?

A. Feed-forward neural networks
B. Recurrent neural networks
C. Convolutional neural networks ✔
D. Reinforcement Learning

Correct answer: C. Convolutional neural networks

11. In which of the following applications can we use deep learning to solve the problem?

A. Prediction of chemical reactions
B. Protein structure prediction
C. Detection of exotic particles
D. All of the above ✔

Correct answer: D. All of the above

12. Which of the following statements is true when you use 1×1 convolutions in a CNN?

A. It suffers less overfitting due to small kernel size
B. It can help in dimensionality reduction
C. It can be used for feature pooling
D. All of the above ✔

Correct answer: D. All of the above

13. Which of the following is well suited for perceptual tasks?

A. Feed-forward Neural Networks
B. Convolutional Neural Networks ✔
C. Reinforcement Learning
D. Recurrent Neural Networks

Correct answer: B. Convolutional Neural Networks

14. Which, if any, of the following propositions is true about fully-connected neural networks (FCNN)?

A. In a FCNN, the most common weight initialization scheme is the zero initialization, because it leads to faster and more robust training.
B. A FCNN with only linear activations is a linear network.
C. In a FCNN, there are connections between neurons of a same layer.
D. None of the above ✔

Correct answer: D. None of the above

15. Which of the following is/are Limitations of Deep Learning?

A. Provide a caption for images
B. Data Labeling
C. Obtain Huge Training Datasets
D. Both C and B ✔

Correct answer: D. Both C and B

16. Which of the following functions can be used as an activation function in the output layer if we wish to predict the probabilities of n classes (p1, p2..pk) such that the sum of p over all n equals to1?

A. ReLu
B. Softmax ✔
C. Sigmoid
D. Tanh

Correct answer: B. Softmax

17. In a neural network, knowing the weight and bias of each neuron is the most important step. If you can somehow get the correct value of weight and bias for each neuron, you can approximate any function. What would be the best way to approach this?

A. Assign random values and pray to God they are correct
B. Iteratively check that after assigning a value how far you are from the best values, and slightly change the assigned values values to make them better ✔
C. Search every possible combination of weights and biases till you get the best value
D. None of these

Correct answer: B. Iteratively check that after assigning a value how far you are from the best values, and slightly change the assigned values values to make them better

18. Which of the following techniques perform similar operations as a dropout in a neural network?

A. Bagging ✔
B. Stacking
C. Boosting
D. None of above

Correct answer: A. Bagging

19. Which of the following would have a constant input in each epoch of training a Deep Learning model?

A. Biases of all hidden layer neurons
B. Activation function of output layer
C. Weight between input and hidden layer ✔
D. Weight between hidden and output layer

Correct answer: C. Weight between input and hidden layer

20. What is a Boltzman machine?

A. A feed forward network with hidden units and probabilistic update
B. A feedback network with hidden units and probabilistic update ✔
C. A feedback network with hidden units
D. A feed forward network with hidden units

Correct answer: B. A feedback network with hidden units and probabilistic update

21. Deep learning is a subset of machine learning that uses:

A. Only linear regression
B. Neural networks with many layers ✔
C. Rule-based systems only
D. Only decision trees

Correct answer: B. Neural networks with many layers

22. An artificial neuron computes:

A. A sorting operation
B. A weighted sum of inputs plus bias, passed through an activation ✔
C. A database query
D. A random number

Correct answer: B. A weighted sum of inputs plus bias, passed through an activation

23. Weights in a neural network are:

A. Learnable parameters that scale inputs ✔
B. Fixed constants
C. Input data
D. Output labels

Correct answer: A. Learnable parameters that scale inputs

24. A bias term allows:

A. Shifting the activation function ✔
B. Removing inputs
C. Normalising data only
D. Fixing weights

Correct answer: A. Shifting the activation function

25. An activation function introduces:

A. Data storage
B. Regularisation only
C. Linearity only
D. Non-linearity into the network ✔

Correct answer: D. Non-linearity into the network

Questions 26 to 50

26. ReLU outputs:

A. Always zero
B. x squared
C. max(0, x) ✔
D. 1 / (1 + e^-x)

Correct answer: C. max(0, x)

27. Sigmoid outputs values between:

A. -1 and 1
B. 0 and 1 ✔
C. -infinity and infinity
D. 0 and infinity

Correct answer: B. 0 and 1

28. Tanh outputs values between:

A. 0 and 1
B. -1 and 1 ✔
C. 0 and 10
D. Any value

Correct answer: B. -1 and 1

29. Softmax is typically used in:

A. Convolution only
B. Input layers
C. Regression outputs only
D. The output layer for multi-class classification ✔

Correct answer: D. The output layer for multi-class classification

30. The vanishing gradient problem makes:

A. Gradients explode
B. Early layers learn very slowly ✔
C. Data disappear
D. Training faster

Correct answer: B. Early layers learn very slowly

31. Exploding gradients can be controlled with:

A. Higher learning rates
B. Removing activations
C. Gradient clipping ✔
D. More layers always

Correct answer: C. Gradient clipping

32. A loss function measures:

A. Model size
B. Number of layers
C. The error between predictions and true values ✔
D. Training time

Correct answer: C. The error between predictions and true values

33. Mean squared error is commonly used for:

A. Object detection only
B. Regression ✔
C. Multi-class classification
D. Clustering

Correct answer: B. Regression

34. Cross-entropy loss is commonly used for:

A. Classification ✔
B. Regression only
C. Dimensionality reduction
D. Data cleaning

Correct answer: A. Classification

35. Backpropagation computes:

A. Learning rates
B. Gradients of the loss with respect to weights ✔
C. Data labels
D. Predictions only

Correct answer: B. Gradients of the loss with respect to weights

36. Gradient descent updates weights:

A. In the direction that increases loss
B. Never
C. In the direction that reduces the loss ✔
D. Randomly

Correct answer: C. In the direction that reduces the loss

37. The learning rate controls:

A. The number of layers
B. The step size of weight updates ✔
C. The dataset size
D. The batch size

Correct answer: B. The step size of weight updates

38. Too high a learning rate can cause:

A. Smaller models
B. Perfect accuracy
C. Very slow learning only
D. Training to diverge or oscillate ✔

Correct answer: D. Training to diverge or oscillate

39. Stochastic gradient descent updates weights using:

A. Small random batches of data ✔
B. No data
C. The entire dataset each step only
D. Only test data

Correct answer: A. Small random batches of data

40. Adam is:

A. An adaptive learning rate optimisation algorithm ✔
B. A dataset
C. A loss function
D. An activation

Correct answer: A. An adaptive learning rate optimisation algorithm

41. An epoch is:

A. One prediction
B. One full pass through the training data ✔
C. One layer
D. One weight update

Correct answer: B. One full pass through the training data

42. Batch size is:

A. The number of layers
B. The number of classes
C. The number of epochs
D. The number of samples processed before a weight update ✔

Correct answer: D. The number of samples processed before a weight update

43. Overfitting in deep learning means:

A. The model memorises training data and generalises poorly ✔
B. Data is balanced
C. The model is too simple
D. Training is too short

Correct answer: A. The model memorises training data and generalises poorly

44. Dropout helps prevent overfitting by:

A. Randomly deactivating neurons during training ✔
B. Adding more neurons
C. Removing the loss
D. Increasing learning rate

Correct answer: A. Randomly deactivating neurons during training

45. L2 regularisation (weight decay) penalises:

A. Large weights ✔
B. Small weights
C. Accuracy
D. Batch size

Correct answer: A. Large weights

46. Early stopping halts training when:

A. Validation performance stops improving ✔
B. Epoch one ends
C. Data runs out
D. Training loss reaches zero always

Correct answer: A. Validation performance stops improving

47. Batch normalisation helps to:

A. Increase overfitting
B. Stabilise and speed up training by normalising layer inputs ✔
C. Remove activations
D. Delete batches

Correct answer: B. Stabilise and speed up training by normalising layer inputs

48. Data augmentation helps by:

A. Creating varied training examples to improve generalisation ✔
B. Reducing data
C. Deleting labels
D. Shrinking models

Correct answer: A. Creating varied training examples to improve generalisation

49. A validation set is used to:

A. Tune hyperparameters and monitor overfitting ✔
B. Train weights directly
C. Store labels
D. Deploy the model

Correct answer: A. Tune hyperparameters and monitor overfitting

50. A CNN is especially suited for:

A. Tabular accounting data only
B. Image and spatial data ✔
C. Text only always
D. Audio only

Correct answer: B. Image and spatial data

Questions 51 to 75

51. A convolutional layer applies:

A. Random noise
B. Learnable filters across the input ✔
C. Fully connected weights only
D. Sorting

Correct answer: B. Learnable filters across the input

52. Pooling layers:

A. Add colour
B. Increase image size
C. Store data
D. Reduce spatial dimensions of feature maps ✔

Correct answer: D. Reduce spatial dimensions of feature maps

53. Max pooling selects:

A. The smallest value
B. The largest value in each region ✔
C. Random values
D. The average value

Correct answer: B. The largest value in each region

54. Stride in convolution is:

A. The filter colour
B. The output class
C. The step size of the filter movement ✔
D. The number of epochs

Correct answer: C. The step size of the filter movement

55. Padding in CNNs is used to:

A. Reduce channels only
B. Add labels
C. Control output size and preserve border information ✔
D. Delete edges

Correct answer: C. Control output size and preserve border information

56. Famous CNN architectures include:

A. LeNet, AlexNet, VGG and ResNet ✔
B. SQL and NoSQL
C. K-means and PCA
D. BERT and GPT only

Correct answer: A. LeNet, AlexNet, VGG and ResNet

57. ResNet introduced:

A. Pooling
B. Dropout
C. Residual (skip) connections ✔
D. Softmax

Correct answer: C. Residual (skip) connections

58. Transfer learning reuses:

A. A pre-trained model for a new related task ✔
B. Test sets
C. Only raw data
D. Random weights

Correct answer: A. A pre-trained model for a new related task

59. Fine-tuning means:

A. Deleting layers
B. Training from scratch only
C. Freezing everything forever
D. Further training a pre-trained model on new data ✔

Correct answer: D. Further training a pre-trained model on new data

60. An RNN is suited for:

A. Tabular data only
B. Sorting
C. Static images only
D. Sequential data such as text or time series ✔

Correct answer: D. Sequential data such as text or time series

61. LSTM networks address:

A. Sorting speed
B. Data storage
C. Image resolution
D. Long-term dependency problems in RNNs ✔

Correct answer: D. Long-term dependency problems in RNNs

62. GRU is:

A. An optimiser
B. A simplified gated recurrent unit ✔
C. A loss function
D. A dataset

Correct answer: B. A simplified gated recurrent unit

63. The Transformer architecture relies on:

A. Decision trees
B. Attention mechanisms ✔
C. Only recurrence
D. Only convolutions

Correct answer: B. Attention mechanisms

64. Self-attention lets a model:

A. Sort tokens
B. Ignore context
C. Delete words
D. Weigh the importance of different tokens in a sequence ✔

Correct answer: D. Weigh the importance of different tokens in a sequence

65. BERT is a Transformer model that is:

A. Bidirectional and pre-trained for language understanding ✔
B. A clustering method
C. A CNN for images
D. An optimiser

Correct answer: A. Bidirectional and pre-trained for language understanding

66. GPT models are:

A. Decision trees
B. Autoregressive Transformer language models ✔
C. Databases
D. CNN classifiers only

Correct answer: B. Autoregressive Transformer language models

67. Tokenisation splits text into:

A. Images
B. Tokens such as words or subwords ✔
C. Numbers only
D. Sentences only

Correct answer: B. Tokens such as words or subwords

68. Word embeddings represent words as:

A. Dense numeric vectors capturing meaning ✔
B. Random strings
C. Database rows
D. One-hot images

Correct answer: A. Dense numeric vectors capturing meaning

69. An autoencoder learns to:

A. Classify only
B. Generate labels
C. Sort data
D. Compress and reconstruct its input ✔

Correct answer: D. Compress and reconstruct its input

70. A GAN consists of:

A. A single loss-free network
B. A generator and a discriminator ✔
C. Two classifiers only
D. An encoder only

Correct answer: B. A generator and a discriminator

71. Diffusion models generate images by:

A. Using decision trees
B. Copying images
C. Learning to reverse a gradual noising process ✔
D. Sorting pixels

Correct answer: C. Learning to reverse a gradual noising process

72. Object detection models like YOLO:

A. Only classify whole images
B. Locate and classify objects in images ✔
C. Only generate text
D. Only sort data

Correct answer: B. Locate and classify objects in images

73. Image segmentation assigns:

A. One label per image
B. A label to each pixel ✔
C. Labels to files only
D. No labels

Correct answer: B. A label to each pixel

74. TensorFlow is:

A. A database
B. A browser
C. An open-source deep learning framework from Google ✔
D. A spreadsheet

Correct answer: C. An open-source deep learning framework from Google

75. PyTorch is:

A. An open-source deep learning framework popular in research ✔
B. An IDE only
C. A web server
D. A database

Correct answer: A. An open-source deep learning framework popular in research

Questions 76 to 100

76. Keras is:

A. A database
B. A high-level API for building neural networks ✔
C. A GPU brand
D. A dataset

Correct answer: B. A high-level API for building neural networks

77. GPUs accelerate deep learning because they:

A. Use less memory always
B. Have more USB ports
C. Store more files
D. Perform many parallel matrix operations ✔

Correct answer: D. Perform many parallel matrix operations

78. A tensor is:

A. A chart
B. A multi-dimensional array ✔
C. A database
D. A single number only

Correct answer: B. A multi-dimensional array

79. Weight initialisation matters because:

A. It has no effect
B. It sets labels
C. It stores data
D. Poor initialisation can slow or prevent learning ✔

Correct answer: D. Poor initialisation can slow or prevent learning

80. Hyperparameters include:

A. Learned weights
B. Learning rate, batch size and number of layers ✔
C. Training labels
D. Predictions

Correct answer: B. Learning rate, batch size and number of layers

81. A confusion matrix evaluates:

A. Model size
B. Training speed
C. Regression error only
D. Classification performance ✔

Correct answer: D. Classification performance

82. Precision and recall are important when:

A. Doing regression only
B. Data is perfectly balanced only
C. Clustering
D. Classes are imbalanced ✔

Correct answer: D. Classes are imbalanced

83. Model quantisation reduces:

A. Accuracy to zero always
B. Model size and inference cost by using lower precision ✔
C. Labels
D. Training data

Correct answer: B. Model size and inference cost by using lower precision

84. Inference means:

A. Collecting data
B. Using a trained model to make predictions ✔
C. Labelling data
D. Training the model

Correct answer: B. Using a trained model to make predictions

85. Explainability tools like Grad-CAM show:

A. Which image regions influenced a prediction ✔
B. Training time
C. Hardware usage
D. Dataset size

Correct answer: A. Which image regions influenced a prediction

86. Bias in deep learning models often comes from:

A. Using Python
B. Using GPUs
C. Large batch sizes
D. Unrepresentative training data ✔

Correct answer: D. Unrepresentative training data

87. A fully connected (dense) layer connects:

A. Every input neuron to every output neuron ✔
B. Only neighbouring pixels
C. Only the first neuron
D. No neurons

Correct answer: A. Every input neuron to every output neuron

88. Flatten layers convert:

A. Text into audio
B. Multi-dimensional feature maps into a vector ✔
C. Labels into data
D. Vectors into images

Correct answer: B. Multi-dimensional feature maps into a vector

89. One-hot encoding of labels represents classes as:

A. Strings
B. Random numbers
C. Binary vectors with a single 1 ✔
D. Images

Correct answer: C. Binary vectors with a single 1

90. Learning rate schedules:

A. Set batch size
B. Adjust the learning rate during training ✔
C. Fix it forever
D. Remove it

Correct answer: B. Adjust the learning rate during training

91. Attention is all you need is the title of the paper introducing:

A. The Transformer ✔
B. ResNet
C. LSTM
D. GANs

Correct answer: A. The Transformer

92. Large language models are trained on:

A. Only audio
B. Only images
C. Only spreadsheets
D. Very large text corpora ✔

Correct answer: D. Very large text corpora

93. Reinforcement learning from human feedback (RLHF) is used to:

A. Align language model outputs with human preferences ✔
B. Store data
C. Compress images
D. Sort tokens

Correct answer: A. Align language model outputs with human preferences

94. Deep learning typically needs:

A. Large amounts of data and compute ✔
B. Only small spreadsheets
C. Only CPUs from 1990
D. No data

Correct answer: A. Large amounts of data and compute

95. Model deployment can use:

A. Only spreadsheets
B. Only paper
C. Only email
D. APIs, mobile runtimes or edge devices ✔

Correct answer: D. APIs, mobile runtimes or edge devices

96. The main strength of deep learning is:

A. Needing no computation
B. Requiring no data
C. Automatically learning complex features from raw data ✔
D. Being fully transparent always

Correct answer: C. Automatically learning complex features from raw data

97. The output layer for binary classification often uses:

A. ReLU only
B. Max pooling
C. No activation always
D. A sigmoid activation ✔

Correct answer: D. A sigmoid activation

98. Leaky ReLU differs from ReLU by:

A. Outputting only ones
B. Always outputting zero
C. Removing gradients
D. Allowing a small gradient for negative inputs ✔

Correct answer: D. Allowing a small gradient for negative inputs

99. Positional encoding in Transformers provides:

A. Learning rate
B. Image colour
C. Batch size
D. Information about token order ✔

Correct answer: D. Information about token order

100. Multi-head attention allows the model to:

A. Attend to information from different representation subspaces ✔
B. Skip attention
C. Delete layers
D. Use only one token

Correct answer: A. Attend to information from different representation subspaces

How to Pass the Deep Learning Exam

  • Read every question carefully; most options look similar, so look for the most precise answer.
  • Revise the questions above topic by topic rather than memorising letters, because option order can change in the live test.
  • You need 80 correct answers out of 100 to pass, so aim to understand at least 85 to stay safe.
  • If you do not clear it on the first attempt, review the questions you missed and use the retake option.

Frequently Asked Questions

How many questions are in the Deep Learning Certification exam?

The exam has 100 multiple-choice questions, each worth 1 mark.

What is the passing score for the Deep Learning Certification?

You need at least 80% (80 out of 100) to pass and receive the certificate.

Can I retake the Deep Learning Certification exam?

Yes. If you score below 80%, you can retry the test after reviewing the topics.

Is the Itronix certificate verifiable?

Yes. Every certificate issued by Itronix Solutions has a unique ID that can be checked on the Itronix verify-certificate page.

More AI & Machine Learning Certification Answers

Disclaimer: These answers are provided for learning and revision. Questions and answers are checked regularly; if you find a mistake, let us know in the comments.

About Priya Dogra

Priya Dogra runs priyadogra.com, where she shares free certification courses, internships, jobs, scholarships and exam preparation guides for Indian students and freshers. Follow her on YouTube, LinkedIn and Instagram for the latest opportunities.

Check Also

Unix/ Linux Certification exam answers 2026

Unix/ Linux Certification Exam Answers 2026 (All 100 Questions)

All 100 Unix/ Linux Certification exam questions with correct answers highlighted - updated 2026.

Leave a Reply

Your email address will not be published. Required fields are marked *