DEEP LEARNING INTERVIEW QUESTIONS

FOR ENTRY LEVEL JOBS

Ideal for students and learners who aspire to enter the fascinating
world of AI, Computer Vision and Deep Learning.

COMPILED BY THE AI EXPERT TEAM AT

Contents

Summary 3
Questions and Answers 3
General Deep Learning 3
Image Classification 21
Object Detection 22
Image Segmentation 23
GANs 26
Miscellaneous 26
CVDL Master Program 36

Deep Learning Interview Questions 2

Summary
Deep Learning is one of the leading and most sought after technologies in the
field of Artificial Intelligence. Deep Learning is widely used in industries like
automotive, healthcare, security, to content creation, the usage of Deep Learning
is exponentially on the rise.

We have compiled this quick reference resource of questions in Deep Learning

and organized them by topics and level of difficulty. This will help you prepare in a
structured manner to successfully clear entry level job interviews. We wish you the
very best in your learning and job preparation journey.

Questions and Answers

General Deep Learning

Question 1 [EASY]: What is Deep Learning, and how does it differ from Machine
Learning?

Answer: Deep learning is a subfield of Machine Learning that focuses on using

artificial neural networks to learn complex patterns and representations from raw
data. While traditional Machine Learning algorithms often require manual feature
extraction and engineering, Deep Learning automatically discovers and extracts
features from the data through the layers of the neural network, enabling it to
handle more complex tasks such as image recognition, natural language
processing, and speech recognition.

Question 2 [EASY]: What is the difference between supervised, unsupervised, and

reinforcement learning?

Answer: Supervised learning is a type of Machine Learning where the model is

trained using labeled data, i.e., input-output pairs. Then, the model learns to map
inputs to the corresponding outputs.
On the other hand, unsupervised learning deals with unlabeled data, and the
model learns to discover patterns or structures in the data without any guidance.

Deep Learning Interview Questions 3

Finally, reinforcement learning is a paradigm where an agent learns to make
decisions by interacting with an environment and receiving feedback in the form
of rewards or penalties to maximize cumulative rewards.

Question 3 [EASY]: What is the role of activation functions in neural networks, and
can you name a few common ones?

Answer: Activation functions are used in neural networks to introduce

non-linearity, enabling the network to learn complex, non-linear patterns in the
data. They are applied to the output of each neuron in the network. Some
common activation functions include:

● Sigmoid

● Hyperbolic Tangent (tanh)

● Rectified Linear Unit (ReLU)

● Leaky ReLU

● Exponential Linear Unit (ELU)

Question 4 [EASY]: Explain the concept of convolutional neural networks (CNNs)

and their applications.

Answer: Convolutional Neural Networks (CNNs) are deep learning architectures

designed explicitly for processing grid-like data, such as images.

CNNs consist of convolutional layers, pooling layers, and fully connected layers.
The convolutional layers apply filters to the input data, capturing local features,
while pooling layers reduce spatial dimensions, helping to control overfitting. As a
result, CNNs have succeeded highly in image classification, object detection, and
semantic segmentation tasks.

Question 5 [EASY]: What is the purpose of loss functions, and can you name a few
common ones?

Deep Learning Interview Questions 4

Answer: Loss, cost, or objective functions measure the difference between the
model's predictions and target values. They guide the training process by
quantifying the model's performance, and the goal is to minimize the loss
function. Some standard loss functions include:

● Mean Squared Error (MSE)

● Cross-Entropy Loss (for classification problems)
● Huber Loss
● Hinge Loss
● Kullback-Leibler Divergence

Question 6 [EASY]: What is the role of optimizers in training neural networks, and
can you name a few common ones?

Answer: Optimizers are algorithms used to update the weights and biases of a
neural network to minimize the loss function. They play a critical role in
determining the speed and effectiveness of model training. Some standard
optimizers include:

● Gradient Descent
● Stochastic Gradient Descent (SGD)
● Momentum
● AdaGrad
● RMSprop
● Adam

Question 7 [EASY]: What are some popular pre-trained deep learning models and
their applications?

Answer: Pre-trained models are deep learning models already trained on large
datasets and can be fine-tuned for specific tasks, leveraging transfer learning.
Some popular pre-trained models and their applications include:

● VGG, ResNet, Inception, and DenseNet for image classification and object
detection
● U-Net for image segmentation

Deep Learning Interview Questions 5

 ● YOLO and Faster R-CNN for real-time object detection
● BERT, GPT, and RoBERTa models for natural language processing tasks, such
as text classification, named entity recognition, and question-answering

Question 8 [EASY]: What is the concept of fine-tuning in the context of transfer

learning?

Answer: Fine-tuning refers to taking a pre-trained deep learning model and

adapting it to a new, related task by updating the model's weights with additional
training on the new task's dataset. This approach leverages the knowledge and
features learned from the original task to achieve better performance and faster
convergence on the new task, especially when the new task has limited labeled
data.

Question 9 [EASY]: How can you apply data augmentation techniques to improve
the performance of a deep learning model?

Answer: Data augmentation involves creating new training samples by applying

various transformations to the original data, such as rotations, translations,
scaling, flipping, and changes in brightness or contrast. This process increases the
diversity and size of the training dataset, helping the model generalize better and
reduce overfitting. In deep learning, data augmentation is typically applied during
training, and the augmented samples are fed into the model along with the
original data.

Question 10 [EASY]: What is transfer learning, and how is it applied in deep

learning for computer vision tasks?

Answer: Transfer learning is a technique in which a pre-trained deep learning

model, usually trained on a large-scale dataset like ImageNet, is fine-tuned or
adapted for a specific, often smaller-scale target task. The rationale is that the
pre-trained model has already learned general features and patterns from the
large dataset, which can be helpful for the target task. In computer vision, transfer
learning is widely used to improve the performance of models in tasks like object
detection, classification, and segmentation, particularly when the target dataset
is small or has limited labeled data.

Deep Learning Interview Questions 6

Question 11 [EASY]: What are some common types of data augmentation used in
deep learning for computer vision tasks?

Answer: Data augmentation is a technique used to increase the diversity and size
of the training dataset by applying random transformations to the input images.
Some common types of data augmentation used in computer vision tasks
include:

● Rotation: Rotating the image by a random angle within a specified range.

● Scaling: Rescaling the image by a random factor within a specified range.
● Flipping: Flipping the image horizontally or vertically with a certain
probability.
● Cropping: Randomly cropping a portion of the image and resizing it to the
original dimensions.
● Translation: Shifting the image randomly along the horizontal and vertical
axes.
● Brightness and contrast adjustment: Modifying the brightness and
contrast of the image by random factors.
● Gaussian noise: Adding random Gaussian noise to the image.
● Color jitter: Randomly changing the image's hue, saturation, and
brightness.

Question 12 [EASY]: What are some popular deep learning frameworks used for
computer vision tasks?

Answer: Several deep learning frameworks have been developed to facilitate

implementing, training, and deploying deep learning models for computer vision
tasks. Some popular frameworks include:

● TensorFlow: An open-source machine learning library developed by

Google with a flexible architecture allowing easy deployment across
multiple platforms.

● Keras: A high-level neural networks API written in Python capable of running

on top of TensorFlow, Microsoft Cognitive Toolkit, or Theano.

Deep Learning Interview Questions 7

 ● PyTorch: An open-source machine learning library developed by Facebook
with a dynamic computation graph and strong support for GPU
acceleration.

● Caffe: A deep learning framework the Berkeley Vision and Learning Center
developed, focusing on image classification and convolutional networks.

● MXNet: A flexible and efficient deep learning library that supports multiple
programming languages and platforms for various machine learning
tasks.

Question 13 [EASY]: What are some challenges and limitations of deep learning for
computer vision tasks?

Answer: Despite the impressive performance of deep learning models in many

computer vision tasks, there are still several challenges and limitations, including:

● Data requirements: Deep learning models often require large amounts of

labeled data for training, which can be expensive and time-consuming.

● Computational resources: Training deep learning models can be

computationally intensive, requiring powerful GPUs or specialized
hardware, which may not be accessible to all users.

● Model interpretability: Deep learning models can be challenging to

interpret, as humans often do not understand the learned features and
decision-making processes.

● Overfitting: Deep learning models, especially those with a large number of

parameters, can be prone to overfitting, leading to poor generalization of
unseen data.

● Adversarial examples: Deep learning models can be sensitive to

adversarial examples, which are carefully crafted input samples designed
to cause the model to make incorrect predictions.

● Domain adaptation: Deep learning models can struggle to generalize to

new or different data distributions, which may require domain adaptation
or transfer learning techniques to address.

Deep Learning Interview Questions 8

Question 14 [MODERATE]: Explain the concept of overfitting and how to prevent it.

Answer: Overfitting occurs when a model learns the noise in the training data
instead of the underlying pattern, resulting in poor performance on unseen data.

There are some standard techniques to prevent overfitting.

● Using more training data

● Reducing model complexity
● Applying regularization techniques like L1 or L2 regularization
● Implementing early stopping
● Using dropout layers in neural networks
● Performing data augmentation

Question 15 [MODERATE]: What is backpropagation, and why is it essential in

training neural networks?

Answer: Backpropagation is an optimization algorithm that trains neural

networks by minimizing errors between actual and predicted outputs. It is based
on the chain rule of calculus and involves computing the gradient of the loss
function with respect to each weight by propagating the gradient back through
the network. The weights and biases get updated using the corresponding
gradients (gradient of loss w.r.t. weight/bias). This process helps the model to
learn and improve its performance over time.

Question 16 [MODERATE]: Explain the concept of batch normalization and its

benefits.

Answer: Batch normalization is a technique used to stabilize and accelerate the

training of deep neural networks. It normalizes the inputs to a layer by adjusting
and scaling their mean and variance, reducing the internal covariate shift. This
process allows for higher learning rates, reduces the sensitivity to weight
initialization, and often leads to faster convergence and better overall
performance.

Deep Learning Interview Questions 9

Question 17 [MODERATE]: What is the purpose of dropout layers in neural
networks?

Answer: Dropout is a regularization technique used in neural networks to prevent

overfitting. During training, dropout layers randomly "drop" a fraction of neurons
by setting their output to zero, making the network more robust and less reliant on
any single neuron. This process encourages the network to learn redundant
representations, effectively simulating the training of multiple smaller networks
and improving generalization performance on unseen data.

Question 18 [MODERATE]: What is the role of weight initialization in training neural

networks, and what are some common strategies?

Answer: Weight initialization plays a crucial role in training neural networks, as it

influences the convergence speed, model performance, and the likelihood of
encountering vanishing or exploding gradients. In addition, proper initialization
can help the network start from a more favorable position in the loss landscape.
Some common weight initialization strategies include:

● Zero initialization
● Random initialization (uniform or normal distribution)
● Xavier/Glorot initialization
● He initialization
● LeCun initialization

Question 19 [MODERATE]: What is early stopping, and how does it help prevent
overfitting in deep learning models?

Answer: Early stopping is a regularization technique to prevent overfitting by

stopping the training process before the model memorizes the noise in the
training data. It involves monitoring a chosen metric, such as validation loss or
accuracy, during training and stopping the training process when the metric
stops improving or starts to degrade. This approach helps to find a balance
between underfitting and overfitting by selecting the model with the best
performance on the validation set.

Deep Learning Interview Questions 10

Question 20 [MODERATE]: What is the role of hyperparameter tuning in deep
learning, and what are some standard methods for it?

Answer: Hyperparameter tuning optimizes the hyperparameters of a deep

learning model, such as learning rate, batch size, number of layers, and number of
neurons per layer, to achieve the best possible performance. Since the optimal
values for these hyperparameters are often problem-specific, a search is
performed to find the best combination. Some standard methods for
hyperparameter tuning include grid search, random search, Bayesian
optimization, and genetic algorithms.

Question 21 [MODERATE]: What are some common challenges in training deep

learning models, and how can they be addressed?

Answer: Some common challenges in training deep learning models include:

● Vanishing and exploding gradients: Addressed by using appropriate

activation functions (e.g., ReLU), weight initialization techniques (e.g., He
initialization), gradient clipping, and skip connections (e.g., ResNet).

● Overfitting: Addressed by regularization techniques (e.g., dropout, L1/L2

regularization), early stopping, data augmentation, and increasing the
training dataset size.

● Underfitting: Addressed by increasing the complexity of the model (e.g.,

adding more layers or neurons), tuning hyperparameters, and using better
optimization algorithms.

● Limited labeled data: Addressed using transfer learning, semi-supervised

learning, data augmentation, or techniques like one-shot or few-shot
learning.

● Computationally expensive training and inference: Hardware accelerators

(e.g., GPUs or TPUs) and distributed training help train the model faster.
Model compression techniques (e.g., pruning, quantization, or knowledge
distillation) help in faster inference.

Deep Learning Interview Questions 11

Question 22 [MODERATE]: Explain the concept of pooling layers in convolutional
neural networks (CNNs) and their benefits.

Answer: Pooling layers are used in CNNs to reduce the spatial dimensions of
feature maps while retaining important information. They apply a downsampling
operation, such as max pooling or average pooling, on non-overlapping regions
of the feature maps. Pooling layers provide several benefits, including:

● Reducing the number of parameters in the model helps prevent overfitting

and lowers computational complexity.
● Introducing translation invariance allows the model to recognize features
regardless of their position in the input.
● Enhancing the model's ability to capture higher-level features by
condensing spatial information.

Question 23 [MODERATE]: How do you handle noisy or missing data in

deep-learning problems?

Answer: Handling noisy or missing data is crucial to ensure deep learning models
can generalize well and provide reliable predictions. Some strategies to deal with
such data include:

● Data cleaning: Removing or correcting noisy instances, outliers, or

duplicate entries in the dataset.

● Data imputation: Filling in missing values with appropriate estimates, such

as the mean, median, or mode of the feature, or using more advanced
techniques like k-Nearest Neighbors imputation or matrix factorization.

● Robust models: Designing models less sensitive to noise or missing values,

such as using dropout layers or incorporating denoising autoencoders.

● Data augmentation: Generating additional training samples with noise to

help the model learn to be more robust to noise during inference.

Deep Learning Interview Questions 12

Question 24 [MODERATE]: What evaluation metrics are commonly used in deep
learning tasks, and what do they measure?

Answer: Evaluation metrics quantify the performance of a deep learning model

and help to compare different models or techniques. Some standard evaluation
metrics include:

● Accuracy: Measures the proportion of correctly classified samples to the

total number of samples. It is commonly used in classification tasks but
can be misleading in cases of class imbalance.

● Precision, Recall, and F1-score: These metrics focus on correctly classifying

positive instances in binary or multi-class classification problems,
considering false positives and false negatives. The F1-score is the
harmonic mean of precision and recall, providing a balanced performance
measure.

● Mean Squared Error (MSE) and Mean Absolute Error (MAE): These metrics
measure the difference between the model's predictions and target values
in regression tasks.

Question 25 [MODERATE]: What is the role of convolutional layers in Convolutional

Neural Networks (CNNs)?

Answer: Convolutional layers are the primary building blocks of CNNs and are
responsible for learning local patterns and features in images. They perform
convolution operations by sliding a set of filters or kernels across the input image,
resulting in feature maps that capture spatial and hierarchical information.
Convolutional layers enable the model to learn features at different scales,
orientations, and positions, providing a robust and translation-invariant
representation for various computer vision tasks.

Question 26 [MODERATE]: What is the purpose of Batch Normalization in deep

learning models, and how does it benefit computer vision tasks?

Answer: Batch Normalization (BN) is a technique used to improve the training of

deep learning models by normalizing each layer's inputs to have zero mean and

Deep Learning Interview Questions 13

unit variance. The BN is achieved by computing the mean and variance of the
mini-batch during training and applying a scaling and shifting operation to
normalize the activations. Batch Normalization helps stabilize and accelerate the
training process, enabling higher learning rates and reducing the sensitivity to
weight initialization. In computer vision tasks, BN is often used in CNNs to improve
convergence speed and generalization performance, leading to better
performance in object detection, classification, and segmentation tasks.

Question 27 [MODERATE]: What are some methods for reducing overfitting in deep
learning models for computer vision?

Answer: Overfitting occurs when a deep learning model learns to perform very
well on the training data but fails to generalize to new, unseen data. Some
methods for reducing overfitting in computer vision models include:

● Data augmentation: Generating additional training samples by applying

random transformations like rotation, scaling, flipping, or cropping to the
input images. These transformations help the model learn to be more
invariant to these transformations and generalize better to unseen data.

● L1/L2 Regularization: Adding regularization terms like L1 or L2 regularization

to the loss function penalizes large model weights and encourages sparsity
or smoothness in the learned features.

● Dropout: Randomly dropping out neurons during training, forcing the

model to learn redundant and more robust representations.

● Early stopping: Monitoring the validation performance during training and

stopping the training process when the validation performance.

Question 28 [MODERATE]: What are some deep learning applications in computer

vision beyond image classification and object detection?

Answer: Deep learning has been successfully applied to a wide range of

computer vision tasks beyond image classification and object detection,
including:

Deep Learning Interview Questions 14

 ● Semantic segmentation: Assigning a class label to each pixel in an image
to create dense, pixel-wise scene labeling.

● Instance segmentation: Labeling each pixel with its corresponding class

while distinguishing between different instances of the same object class.

● Image synthesis and generation: Generating realistic images, either from

random noise (e.g., GANs) or based on a given input (e.g., style transfer,
image-to-image translation).

● Image super-resolution: Upsampling low-resolution images to higher

resolutions while preserving or enhancing image details.

● Optical character recognition (OCR): Recognizing and extracting text from

images.

● 3D object recognition and reconstruction: Identifying objects in 3D space

and reconstructing their shapes and poses from 2D images or depth data.

● Action recognition and video analysis: Classifying actions or events in

video sequences and understanding temporal dependencies in the data.

● Image captioning: Generating textual descriptions of images based on

their content.

Question 29 [MODERATE]: What is multi-task learning, and how is it applied in

deep learning for computer vision tasks?

Answer: Multi-task learning is a training strategy where a single deep learning

model is trained to perform multiple related tasks simultaneously, sharing the
learned features and representations between the tasks. The rationale behind
multi-task learning is that learning multiple related tasks can lead to better
generalization and improved performance, as the model can exploit the shared
structure and underlying patterns across the tasks.

In computer vision, multi-task learning is often applied by having a shared

backbone network (e.g., a CNN) for feature extraction and multiple task-specific
heads or branches for different tasks, such as object detection, classification, and
segmentation. The model is trained on a combined loss function, the sum or

Deep Learning Interview Questions 15

weighted sum of the individual task losses. Multi-task learning has been used in
various computer vision applications, such as autonomous driving, where the
model must simultaneously perform tasks like object detection, lane detection,
and depth estimation.

Question 30 [MODERATE]: How does active learning help reduce the labeling effort
in deep learning for computer vision tasks?

Answer: Active learning is a semi-supervised learning approach in which the

model actively selects the most informative and uncertain samples from the
unlabeled data for human annotation to improve its performance using as few
labeled samples as possible. By prioritizing the samples likely to impact the
model's learning most, active learning can reduce the labeled data required for
training, minimizing the labeling effort and cost.

In computer vision tasks, active learning can be applied by training an initial

model on a small labeled dataset and then using the model to predict the labels
for a larger unlabeled dataset. The model's uncertainty or informativeness can be
measured using criteria like entropy, least confidence, or margin sampling. The
most uncertain or informative samples are then selected for human annotation,
and the model is fine-tuned on the updated labeled dataset. This process can be
repeated iteratively to improve the model's performance progressively.

Question 31 [MODERATE]: What are some methods for improving the efficiency of
deep learning models for computer vision tasks, especially for deployment on
resource-constrained devices?

Answer: Several methods can be used to improve the efficiency of deep learning
models for computer vision tasks, particularly when deploying on
resource-constrained devices like mobile phones or embedded systems. Some
popular methods include:

● Model compression: Techniques like weight pruning, quantization, and

knowledge distillation can reduce the model size and computational
complexity while maintaining similar performance.

Deep Learning Interview Questions 16

 ● Network architecture design: Designing more efficient network
architectures, such as MobileNet or SqueezeNet, which use depth-wise
separable convolutions or other techniques to reduce the number of
parameters and computations.

● Model Optimization: Applying software optimization techniques, such as

graph optimization, kernel fusion, and hardware-specific optimizations, to
speed up the model's inference time.

● Hardware acceleration: Utilizing specialized hardware, like GPUs, TPUs, or

custom accelerators, to accelerate the computation of deep learning
models.

Question 32 [MODERATE]: What are some methods for handling class imbalance
in deep learning for computer vision tasks?

Answer: Class imbalance is a common issue in many computer vision tasks,

where some classes have significantly fewer examples than others in the training
data. This can lead to poor performance in the underrepresented classes, as the
model may be biased towards the majority classes. Some methods for handling
class imbalance in deep learning for computer vision tasks include:

● Data augmentation: Creating more training examples for the minority

classes by applying various transformations, such as rotation, scaling, or
flipping, to the original images.

● Resampling: Oversampling the minority classes, undersampling the

majority classes, or combining both to create a more balanced training
dataset.

● Loss function modification: Adjusting the loss function to assign higher

importance or penalties to the minority classes, such as using
class-weighted cross-entropy loss or focal loss.

● Transfer learning: Pretraining the model on a larger, more balanced

dataset, and then fine-tuning it on the imbalanced target dataset, which
can help the model learn more generalizable features.

Deep Learning Interview Questions 17

 ● Ensemble methods: Combining the predictions of multiple models trained
on different subsets of the data, which can help improve the overall
performance and reduce the bias towards majority classes.

Question 33 [MODERATE]: What is the role of synthetic data in deep learning for
computer vision tasks?

Answer: Synthetic data refers to artificially generated data created using

computer graphics, simulations, or other techniques rather than being collected
from real-world sources. Synthetic data can be crucial in deep learning for
computer vision tasks, particularly when obtaining labeled data is difficult,
expensive, or time-consuming.

In deep learning for computer vision, synthetic data can be used for various
purposes, such as:

● Data augmentation: Generating additional training examples to increase

the size and diversity of the training dataset can help improve the model's
performance and generalization.

● Domain adaptation: Creating synthetic data that simulates the target

domain, allowing the model to adapt more effectively to a new
environment or application.

● Rare event simulation: Generating synthetic examples of rare events or

situations that may be difficult or impossible to capture in real-world data,
such as extreme weather conditions, accidents, or unusual object
configurations.

● Privacy preservation: Creating synthetic data that maintains the statistical

properties of the original data while removing personally identifiable
information, which can help address privacy concerns in sensitive
applications.

While synthetic data can be a valuable resource in deep learning for computer
vision tasks, ensuring that the generated data represents real-world data and
that the model is balanced with the synthetic data's specific characteristics is
essential. Techniques like domain randomization or mixing synthetic and real

Deep Learning Interview Questions 18

data during training can help mitigate these issues and improve the model's
generalization to real-world scenarios.

Question 34 [DIFFICULT]: Explain the concept of skip connections and their

benefits in deep learning architectures.

Answer: Skip connections, shortcuts, or residual connections allow information to

bypass one or more layers in a deep learning model. They are a key component
of architectures like ResNet and DenseNet. Skip connections help address the
vanishing gradient problem by allowing gradients to flow more easily through the
network during backpropagation. They also enable the construction of deeper
models without sacrificing performance, as the added layers can learn residual
mappings instead of full mappings.

Question 35 [DIFFICULT]: What is the difference between feature-based and

appearance-based methods in computer vision?

Answer: Feature-based methods in computer vision rely on extracting distinctive,

invariant features from images and using them to perform tasks like object
recognition or matching. Examples of feature-based methods include SIFT, SURF,
and ORB. On the other hand, appearance-based methods involve learning
models that directly use the raw pixel values of images or their spatial
representations, such as convolutional neural networks (CNNs), to perform tasks
like object detection and segmentation or classification.

Question 36 [DIFFICULT]: What are some ethical considerations in developing and

deploying deep learning models for computer vision tasks?

Answer: The development and deployment of deep learning models for computer
vision tasks come with several ethical considerations, including:

● Bias and fairness: Models can learn biases in the training data, leading to
unfair treatment of certain groups or individuals. Ensuring that the training
data is representative and diverse and that the model's performance is
evaluated across different demographic groups is crucial.

Deep Learning Interview Questions 19

 ● Privacy: Computer vision models, particularly those used for surveillance or
facial recognition, can infringe on individual privacy rights. Developers
should consider implementing privacy-preserving techniques like
differential privacy or federated learning, and organizations should
establish clear policies on data collection and usage.

● Transparency and interpretability: Deep learning models can be difficult to

interpret, making it challenging to explain their decisions or identify
potential issues. Developers should strive for model interpretability and
document the model's behavior and limitations.

● Accountability and responsibility: Deploying computer vision models in

sensitive applications, such as autonomous vehicles or medical diagnosis,
can have significant consequences. Establishing clear lines of
accountability and responsibility for the developers, users, and
organizations involved in developing and deploying these models is
essential.

● Environmental impact: Training deep learning models can be

computationally intensive, consuming large amounts of energy and
contributing to carbon emissions. Developers should consider the
environmental impact of their models and explore techniques to reduce
energy consumption, such as model compression or more efficient
architectures.

Question 37 [DIFFICULT]: What are some standard techniques for model

compression in deep learning for computer vision tasks?

Answer: Model compression refers to reducing a deep learning model's size,

memory footprint, or computational complexity while maintaining its
performance. Model compression is essential for deploying deep learning models
on resource-constrained devices, such as mobile phones or edge devices, where
memory and computation resources are limited. Some standard techniques for
model compression in deep learning for computer vision tasks include:

Deep Learning Interview Questions 20

 ● Pruning: Removing redundant or less important connections or neurons
from the model reduces the parameters and computations required for
inference.

● Quantization: Reducing the precision of the model's weights and

activations, such as using lower-precision floating-point or integer
representations, can significantly reduce the model's memory footprint and
computational cost.

● Knowledge distillation: Training a smaller, more compact model (student)

to mimic the behavior of a larger, more complex model (teacher), allowing
the student model to achieve similar performance with fewer parameters
and computations.

● Model architecture search: Exploring different model architectures, layer

types, or connectivity patterns to find more efficient and compact models
that maintain the desired performance.

● Weight sharing: Using the same weights for multiple connections in the
model can reduce the number of unique parameters and memory
requirements.

Image Classification
Question 38 [MODERATE]: How do you handle a class imbalance in deep learning
problems?

Answer: Class imbalance occurs when certain classes in the dataset have
significantly fewer samples than others, leading to biased model predictions.
Some techniques to handle class imbalance include:

● Resampling the data (oversampling the minority class or undersampling

the majority class)

● Using data augmentation to generate more samples for the minority class

● Applying cost-sensitive learning by assigning different weights to classes

Deep Learning Interview Questions 21

 ● Using evaluation metrics like F1-score, precision, recall, or the area under
the ROC curve (AUC-ROC) that are less sensitive to class imbalance.

Object Detection
Question 39 [EASY]: What is the difference between object detection and object
recognition in computer vision?

Answer: Object recognition is the process of identifying the class or category of

an object in an image. In contrast, object detection involves recognizing and
localizing the object within the image by providing a bounding box around it.
Object detection is a more challenging task as it requires the model to handle
various scales, orientations, and occlusions of objects in the image.

Question 40 [EASY]: What are some popular object detection algorithms used in
computer vision?

Answer: Some popular object detection algorithms include:

● R-CNN and its variants (Fast R-CNN, Faster R-CNN)

● YOLO (You Only Look Once) and its variants (YOLOv2, YOLOv3, YOLOv4,
YOLOv5)

● SSD (Single Shot MultiBox Detector)

● RetinaNet
● EfficientDet

Question 41 [EASY]: What are anchor boxes in object detection, and what is their
purpose?

Answer: Anchor boxes, priors, or default boxes are pre-defined bounding boxes of
different shapes, sizes, and aspect ratios used in object detection algorithms like
YOLO and SSD. They aim to provide a set of initial reference boxes that adjust
during training to match the ground truth bounding boxes more closely. Anchor
boxes help the model learn to predict bounding boxes more efficiently by

Deep Learning Interview Questions 22

providing a starting point close to the correct scale and aspect ratio of the
objects in the dataset.

Question 42 [MODERATE]: Explain the concept of non-maximum suppression

(NMS) in object detection.

Answer: Non-maximum suppression (NMS) is a post-processing step used in

object detection algorithms to eliminate overlapping or redundant bounding
boxes and select the most accurate ones. NMS works by considering the detection
confidence scores of all predicted bounding boxes and selecting the one with the
highest score. It then suppresses any other bounding boxes with a significant
overlap (measured using Intersection over Union or IoU) with the chosen box. This
process is repeated until all remaining boxes have been either selected or
suppressed.

Image Segmentation
Question 43 [EASY]: What is the difference between semantic segmentation and
instance segmentation in computer vision?

Answer: Semantic segmentation is a computer vision task where each pixel in an

image is assigned a class label, indicating the object category it belongs to. It
focuses on identifying the boundaries and regions of different object classes in
the image but does not distinguish between individual instances of the same
class. Instance segmentation, on the other hand, is a more challenging task that
involves not only classifying each pixel but also separating different instances of
the same object class, providing both the object category and instance
information.

Question 44 [MODERATE]: What are dilated convolutions, and how can they help
capture larger contextual information in deep learning models?

Answer: Dilated or atrous convolutions are a variant of standard convolutions

that incorporate a dilation factor, which controls the spacing between the filter's
weights. By increasing the dilation factor, the receptive field of the convolution
operation is expanded, allowing the model to capture larger contextual

Deep Learning Interview Questions 23

information without increasing the number of parameters or computational
complexity. Dilated convolutions are particularly useful in tasks like semantic
segmentation, where capturing long-range dependencies and multi-scale
information is crucial for accurate pixel-wise predictions.

Question 45 [MODERATE]: What is a Fully Convolutional Network (FCN), and how is

it used for semantic segmentation?

Answer: A Fully Convolutional Network (FCN) is a convolutional neural network

designed for pixel-wise dense prediction tasks like semantic segmentation. Unlike
traditional CNNs that use fully connected layers for classification, FCNs replace
these fully connected layers with convolutional layers to produce spatially dense
outputs that allow the network to accept input images of any size and generate
segmentation masks with the same spatial dimensions. In addition, FCNs often
employ upsampling layers or transposed convolutions to recover the resolution of
the input image, resulting in pixel-wise predictions.

Question 46 [MODERATE]: What are Region Proposal Networks (RPNs) in the

context of object detection algorithms?

Answer: Region Proposal Networks (RPNs) are a key Faster R-CNN object detection
algorithm component. They generate a set of candidate object bounding boxes
or region proposals, which are then passed to a classifier and regressor to predict
the object class and refine the bounding box coordinates. RPNs are fully
convolutional networks that leverage the feature maps of a CNN backbone to
efficiently generate region proposals in a sliding window fashion, significantly
speeding up the object detection process compared to earlier methods like
R-CNN and Fast R-CNN.

Question 47 [MODERATE]: What are U-Net and Mask R-CNN, and how do their
semantic and instance segmentation approaches differ?

Answer: U-Net is a deep learning architecture designed explicitly for semantic

segmentation. It has an encoder-decoder structure with skip connections
between corresponding encoder and decoder layers, allowing the model to
capture high-level and low-level features and produce accurate segmentation

Deep Learning Interview Questions 24

masks. U-Net is mainly used for semantic segmentation tasks, where the goal is
to assign a class label to each pixel in an image.

On the other hand, Mask R-CNN is an extension of the Faster R-CNN object
detection algorithm, designed for instance segmentation. It adds a branch to the
Faster R-CNN architecture to predict binary masks for each object instance, class
labels, and bounding box coordinates. Mask R-CNN is used in instance
segmentation tasks, where the goal is to label each pixel with its corresponding
class and distinguish between different instances of the same object class.

Question 48 [DIFFICULT]: What are skip connections in deep learning models, and
how do they improve performance in segmentation tasks?

Answer: Skip connections are a technique used in deep learning models to

connect the output of a layer to the input of a non-adjacent layer, creating a
shortcut or bypass around intermediate layers. They help mitigate the vanishing
gradient problem in deep networks, enabling more effective training of deeper
models. In computer vision tasks, skip connections are often used in
encoder-decoder architectures, such as U-Net and ResNet, to preserve spatial
information from early layers and improve the output quality, particularly in tasks
like semantic segmentation and image super-resolution.

Question 49 [DIFFICULT]: What is the spatial pyramid pooling (SPP) concept in

CNNs, and how does it benefit computer vision tasks?

Answer: Spatial pyramid pooling (SPP) is a technique used in CNNs to enable the
processing of input images of varying sizes and aspect ratios. SPP divides the
feature map generated by the convolutional layers into a fixed number of
sub-regions at different scales, pooling the features within each sub-region. The
pooled features are then concatenated to form a fixed-length representation,
which can be fed into the subsequent fully connected layers or classifiers. SPP
allows CNNs to handle images of different sizes without resizing or cropping,
improving performance in object detection and recognition tasks.

Deep Learning Interview Questions 25

GANs
Question 50 [EASY]: What is a Generative Adversarial Network (GAN), and how
does it work?

Answer: A Generative Adversarial Network (GAN) is a deep learning architecture

comprising two neural networks, a generator, and a discriminator, trained in a
zero-sum game (adversarial) setting. The generator learns to create synthetic
data that resembles the real data distribution, while the discriminator learns to
distinguish between real and generated samples. The training process involves
both networks improving their capabilities iteratively, resulting in the generator
producing increasingly realistic samples.

Miscellaneous
Question 51 [EASY]: What is optical flow, and how is it used in computer vision?

Answer: Optical flow is the apparent motion of objects, surfaces, and edges in a
sequence of images or video frames caused by the relative motion between the
camera and the scene. It is used in computer vision tasks such as motion
estimation, object tracking, and video stabilization. Optical flow algorithms
estimate the motion vectors for each pixel in the image, which can be used to
understand the dynamics of the scene and predict the future positions of objects
or points of interest.

Question 52 [MODERATE]: What are Siamese networks, and how are they used in
computer vision tasks?

Answer: Siamese networks are neural network architectures designed to compare

or measure the similarity between two input samples. They consist of two
identical sub-networks that share the same weights and are trained
simultaneously. The outputs of the two sub-networks are combined using a
distance metric or similarity function to produce a single scalar value,
representing the similarity or dissimilarity between the input samples. Siamese

Deep Learning Interview Questions 26

networks are used in computer vision tasks such as face verification, signature
verification, and one-shot learning.

Question 53 [EASY]: Explain the concept of Recurrent Neural Networks (RNNs) and
their applications.

Answer: Recurrent Neural Networks (RNNs) are neural network architectures

designed to handle sequence data or data with temporal dependencies. They
contain loops that allow information to persist across time steps, making them
suitable for time series forecasting, natural language processing, and speech
recognition.

RNNs suffer from vanishing and exploding gradient problems, which have led to
the development of more advanced architectures like LSTM and GRU.

Question 54 [EASY]: What are Long Short-Term Memory (LSTM) networks, and how
do they differ from regular RNNs?

Answer: Long Short-Term Memory (LSTM) networks are a particular type of RNN
designed to address the vanishing and exploding gradient problems that hinder
the training of standard RNNs. LSTMs introduce a memory cell and three gating
mechanisms (input, output, and forget gates) that regulate the flow of
information within the cell. This design allows LSTMs to learn and remember
long-range dependencies more effectively than regular RNNs, making them
suitable for tasks like machine translation, text generation, and sentiment
analysis.

Question 55 [EASY]: What is the difference between an autoencoder and a

variational autoencoder?

Answer: An autoencoder is an unsupervised neural network architecture for data

compression, denoising, and representation learning. It consists of an encoder
that compresses the input data into a lower-dimensional representation (latent
space) and a decoder that reconstructs the input data from the latent
representation. A variational autoencoder (VAE) is a generative variant of an

Deep Learning Interview Questions 27

autoencoder, which adds a probabilistic layer to model the latent space
distribution. VAEs enable the generation of new samples by sampling from the
latent space distribution.

Question 56 [EASY]: What is the difference between unsupervised, supervised, and

self-supervised learning in deep learning for computer vision?

Answer: In supervised learning, models are trained using labeled data, where
each input sample is associated with a corresponding output label or target.
Supervised learning is the most common approach in deep learning for computer
vision tasks, such as image classification, object detection, and semantic
segmentation.

On the other hand, unsupervised learning involves training models using only the
input data without any corresponding output labels. The goal is to learn the
underlying structure or patterns in the data. Unsupervised learning methods, such
as clustering and dimensionality reduction, can be applied in computer vision
tasks like image segmentation, feature learning, and data compression.

Self-supervised learning is a type of unsupervised learning where models are

trained using labels automatically generated from the input data without human
annotation. This approach leverages the structure and inherent patterns in the
data to create auxiliary tasks that can be used to learn useful representations. For
example, in computer vision, self-supervised learning techniques include
contrastive learning, where models learn to distinguish between similar and
dissimilar image patches, and pretext tasks, such as predicting the relative
position of image patches or solving jigsaw puzzles.

Question 57 [MODERATE]: Explain the concept of attention mechanisms in deep

learning.

Answer: Attention mechanisms are used in deep learning models, particularly in

sequence-to-sequence tasks, to allow the model to focus on relevant input parts
when making predictions. By assigning different weights to different input parts,
attention mechanisms help models capture long-range dependencies and
improve performance in tasks such as machine translation, text summarization,
and image captioning.

Deep Learning Interview Questions 28

Question 58 [MODERATE]: What is the difference between a one-shot learning and
a few-shot learning problem?

Answer: One-shot and few-shot learning are both types of learning problems that
deal with small amounts of labeled data.

In one-shot learning, the model must learn to recognize new objects or classes
based on just one or very few examples, while few-shot learning involves learning
from a slightly larger but still limited number of examples, usually less than ten
per class. Both paradigms require models to generalize effectively from scarce
data and often rely on techniques like transfer learning, meta-learning, or
memory-augmented neural networks.

Question 59 [MODERATE]: What is the teacher-student paradigm in deep learning,

and how does it work?

Answer: The teacher-student paradigm, also known as knowledge distillation, is

a technique used to transfer the knowledge learned by a larger, more complex
model (teacher) to a smaller, more efficient model (student). In this training
process, the student model mimics the teacher model's output probabilities or
intermediate representations, resulting in a smaller model that performs similarly
to the larger one. This approach helps deploy deep learning models on
resource-constrained devices or reduce inference time.

Question 60 [MODERATE]: What are word embeddings, and how are they used in
natural language processing (NLP) tasks?

Answer: Word embeddings are continuous vector representations of words that

capture their semantic meaning and relationships with other words. They are
used in NLP tasks to convert discrete text data into continuous, fixed-size vectors
that neural networks can process. Word embeddings are learned from large text
corpora using algorithms like Word2Vec, GloVe, or FastText, and can be fine-tuned
for specific tasks. The word embeddings enable models to understand the
similarities and relationships between words and improve performance in tasks
such as text classification, sentiment analysis, and machine translation.

Deep Learning Interview Questions 29

Question 61 [MODERATE]: What is the Transformer architecture, and how has it
impacted natural language processing?

Answer: The Transformer architecture, introduced in the paper "Attention is All You
Need" by Vaswani et al., is a deep learning model that relies solely on
self-attention mechanisms instead of traditional recurrent or convolutional layers.
It has significantly impacted NLP due to its ability to capture long-range
dependencies more effectively and its highly parallelizable structure, which
enables faster training. As a result, the Transformer has become the foundation
for many state-of-the-art models like BERT, GPT, and T5, which have advanced
the performance of various NLP tasks, including machine translation, text
classification, and question-answering.

Question 62 [MODERATE]: What is a seq2seq model, and what are its applications?

Answer: A seq2seq (sequence-to-sequence) model is a type of deep learning

architecture designed to handle tasks that involve mapping input sequences to
output sequences. It typically consists of an encoder network that processes the
input sequence and a decoder network that generates the output sequence.
Seq2seq models are often used in machine translation, text summarization, and
conversational agent tasks. Advanced variants of seq2seq models incorporate
attention mechanisms to improve their ability to capture long-range
dependencies and handle more complex tasks.

Question 63 [MODERATE]: What is an image captioning task, and what are some
deep learning approaches to solve it?

Answer: Image captioning is a multi-modal task that involves generating a

natural language description of the content of an image. It requires a model to
understand visual and textual information and establish relationships. Some
deep-learning approaches to image captioning include:

● Encoder-decoder architectures: Combining a CNN as an encoder to

extract visual features from the image with an RNN or Transformer-based
decoder to generate the caption.

● Attention mechanisms: Incorporating attention layers allows the model to

focus on relevant parts of the image when generating each word in the
caption, improving the text's quality and coherence.

Deep Learning Interview Questions 30

Question 64 [MODERATE]: What is the role of recurrent neural networks (RNNs) in
computer vision tasks?

Answer: Recurrent Neural Networks (RNNs) are a type of neural network

architecture designed to handle sequential data by maintaining a hidden state
that can capture information from previous time steps. In computer vision tasks,
RNNs are often combined with CNNs to model temporal dependencies in video
data, such as in video classification, action recognition, and object tracking. RNNs
can also be used for image captioning. They generate a textual description of an
input image by conditioning the language model on the visual features extracted
by a CNN.

Question 65 [MODERATE]: What are the key differences between 2D and 3D

convolutional neural networks?

Answer: The main difference between 2D and 3D convolutional neural networks

(CNNs) lies in the dimensions of the input data and the convolution operations.
While 2D CNNs are designed to process 2D images (e.g., height and width) with 2D
convolution operations, 3D CNNs are built to handle 3D volumetric data (e.g.,
height, width, and depth) using 3D convolution operations. In a 3D CNN, the filters
or kernels have an additional depth dimension, and the convolution operation is
performed across all three spatial dimensions. 3D CNNs are commonly used for
tasks involving 3D data, such as medical image analysis, 3D object recognition,
and video analysis.

Question 66 [MODERATE]: How do attention mechanisms help improve the

performance of deep learning models in computer vision tasks?

Answer: Attention mechanisms in deep learning models allow the model to

selectively focus on specific regions or features in the input data, dynamically
weighting their importance based on the task and context. By incorporating
attention, the model can learn to prioritize relevant information and ignore
irrelevant or noisy parts of the data. In computer vision tasks, attention
mechanisms have been applied in various ways, such as spatial attention to
focus on specific regions in an image, channel-wise attention to emphasize
certain feature channels, and temporal attention to focus on specific time steps in
video sequences. Attention mechanisms have been shown to improve the

Deep Learning Interview Questions 31

performance of deep learning models in tasks like image captioning, object
detection, and video analysis.

Question 67 [MODERATE]: What is the role of reinforcement learning in deep

learning for computer vision tasks?

Answer: Reinforcement learning (RL) is a type of machine learning in which an

agent learns to make decisions by interacting with an environment and receiving
feedback in the form of rewards or penalties. While RL is not a primary method for
most computer vision tasks, it can be combined with deep learning techniques to
address specific problems, such as visual navigation, robotic manipulation, or
active object recognition.

In deep learning for computer vision, RL can be applied using a deep neural
network, such as a CNN, as a function approximator for the agent's policy or value
function. The network can be trained using RL algorithms like Q-learning or policy
gradients to learn the optimal actions based on the visual input from the
environment. This combination of deep learning and reinforcement learning,
known as deep reinforcement learning (DRL), has succeeded in several computer
vision-related tasks, including self-driving cars, robot control, and video game
playing.

Question 68 [MODERATE]: What is the role of Transformers in deep learning for

computer vision tasks?

Answer: Transformers are a type of neural network architecture initially designed

for natural language processing tasks, known for their self-attention mechanisms
and ability to model long-range dependencies. Recently, Transformers have been
adapted and applied to computer vision tasks, demonstrating impressive
performance in various areas such as image classification, object detection, and
segmentation.

In deep learning for computer vision, Transformers can be used as an alternative

or complement to traditional CNNs. Unlike CNNs, which use local convolutions to
capture spatial information, Transformers can model global dependencies
between pixels through self-attention mechanisms. The self-attention allows
them to learn and exploit long-range relationships in the input data more
effectively.

Deep Learning Interview Questions 32

Some popular computer vision models based on Transformers include Vision
Transformers (ViT), DETR (DEtection TRansformer), and Swin Transformers. These
models have achieved state-of-the-art performance on various benchmarks,
indicating that Transformers have significant potential for future developments in
computer vision.

Question 69 [MODERATE]: What is the role of capsule networks in deep learning for
computer vision tasks?

Answer: Capsule networks are a type of neural network architecture proposed by

Geoffrey Hinton as an alternative to traditional CNNs for computer vision tasks.
The key idea behind capsule networks is using "capsules," which are small groups
of neurons representing different aspects of an object, such as its pose, scale, or
orientation. These capsules are designed to capture the hierarchical relationships
between different parts of an object and maintain spatial information throughout
the network.

In deep learning for computer vision tasks, capsule networks aim to address
some of the limitations of CNNs, such as the lack of explicit spatial relationships
between features and the difficulty in modeling viewpoint invariance. Capsule
networks have shown promising results in tasks like object recognition and
segmentation, but their performance and scalability are still active research
areas. Nevertheless, further development of capsule networks may lead to more
robust and interpretable models for computer vision tasks.

Question 70 [MODERATE]: What is the role of graph neural networks (GNNs) in

deep learning for computer vision tasks?

Answer: Graph neural networks (GNNs) are a type of neural network architecture
designed to process graph-structured data, which consists of nodes connected
by edges representing relationships between them. While GNNs are primarily used
in domains like social network analysis, recommendation systems, and molecular
chemistry, they can also be applied to specific computer vision tasks that involve
non-grid structured data or complex relationships between objects.

In deep learning for computer vision, GNNs can be used to model the
relationships between different regions or objects in an image or a video. For
example, in scene understanding tasks, GNNs can capture the relationships

Deep Learning Interview Questions 33

between different objects and their attributes, such as "person riding a bike" or
"cat sitting on a chair." Similarly, GNNs can model the interactions between
objects over time in video analysis tasks, such as tracking objects in a video or
understanding the relationships between objects and their trajectories.

Question 71 [MODERATE]: How can adversarial training improve the robustness of

deep learning models in computer vision tasks?

Answer: Adversarial training is a technique that involves generating adversarial

examples, which are modified input samples designed to fool the model, and
using them as additional training data. By exposing the model to these
adversarial examples during training, the model learns to recognize and defend
against such attacks, improving robustness and generalization.

In deep learning for computer vision tasks, adversarial training can be applied by
generating adversarial examples using the Fast Gradient Sign Method (FGSM) or
Projected Gradient Descent (PGD), which perturb the input images in a way that
maximizes the model's prediction error. These adversarial examples are then
mixed with the original training data and used to update the model's weights.
Adversarial training has been shown to improve the model's robustness against
adversarial attacks and can also lead to better performance on clean,
non-adversarial data in some cases.

Question 72 [DIFFICULT]: How can unsupervised learning be applied in deep

learning for computer vision tasks?

Answer: Unsupervised learning is a type of machine learning in which a model

learns to discover patterns or structures in the data without using labeled
examples. For example, unsupervised learning can be applied to feature learning,
clustering, and representation learning tasks in deep learning for computer vision.

Some common unsupervised learning techniques used in computer vision

include:

● Autoencoders: Neural networks that learn to encode and decode the input
data, forcing the model to learn a compact and useful representation of
the data in the hidden layers.

Deep Learning Interview Questions 34

 ● Generative models: Models like Generative Adversarial Networks (GANs)
and Variational Autoencoders (VAEs) that learn to generate realistic
samples from the data distribution, which can be used for tasks like image
synthesis, inpainting, or style transfer.

● Clustering: Unsupervised methods like k-means or hierarchical clustering

can be applied to the feature space learned by deep learning models to
group similar images or discover underlying structures in the data.

● Self-supervised learning: Techniques that generate "pseudo-labels" from

the input data, such as predicting the relative position of image patches or
solving jigsaw puzzles, which can be used to pre-train the model before
fine-tuning on a supervised task.

Question 73 [DIFFICULT]: What are some techniques for incorporating spatial and
temporal information in deep learning models for computer vision tasks?

Answer: Spatial and temporal information is essential for many computer vision
tasks, particularly those involving video analysis or sequences of images. Some
techniques for incorporating spatial and temporal information in deep learning
models for computer vision tasks include:

● 3D Convolutional Neural Networks (3D CNNs): Extending traditional 2D

CNNs with an additional spatial dimension, allowing the model to process
volumetric data or sequences of images more effectively.

● Convolutional LSTM (ConvLSTM): Combining convolutional operations with

Long Short-Term Memory (LSTM) cells to capture spatial and temporal
dependencies in the input data.

● Recurrent Neural Networks (RNNs): Using RNNs, such as LSTMs or GRUs, to

model the temporal relationships between frames in a video or the output
of a CNN feature extractor.

● Temporal pooling: Aggregating the features from multiple frames or time

steps using pooling operations, such as max pooling or average pooling, to
capture the temporal information.

Deep Learning Interview Questions 35

 ● Attention mechanisms: Incorporating attention mechanisms, such as
self-attention or temporal attention, selectively focus on specific regions or
time steps in the input data, allowing the model to capture long-range
dependencies more effectively.

CVDL Master Program

Become an OpenCV certified AI Professional with our CVDL Master Program. This
flagship program from OpenCV University is the world’s most comprehensive
curation of beginner to expert level courses in Computer Vision, Deep Learning,
and AI.

By taking the program, you become part of CareerX, our Career Accelerator
Program curated to help you progress your career as an AI Professional.

For any questions, please drop an email at university@opencv.org.

Deep Learning Interview Questions 36