Certified Nephrology Nurse Exam — Questions and Answers

Question 1: During the training of a deep CNN, a practitioner observes that the distribution of inputs to deeper layers is constantly changing, a phenomenon known as internal covariate shift. This slows down training because the layers must continually adapt to a new distribution. Which technique is specifically designed to mitigate this problem by normalizing the inputs to each layer?

Gradient Clipping
L2 Regularization
Learning Rate Annealing
Batch Normalization (Correct answer)

Correct answer: Batch Normalization

Batch Normalization is a technique designed to reduce internal covariate shift. [11] It normalizes the output of a previous activation layer by subtracting the batch mean and dividing by the batch standard deviation. [9, 11] This stabilization of the distributions of layer inputs allows for faster training, higher learning rates, and can act as a form of regularization. [10, 11]

Question 2: Which popular pretrained CNN architecture introduced depthwise separable convolutions, making it efficient for transfer learning on mobile devices?

VGG-16
ResNet-50
MobileNet (Correct answer)
AlexNet

Correct answer: MobileNet

MobileNet uses depthwise separable convolutions to dramatically reduce computation, making it ideal for mobile and embedded transfer learning.

Question 3: What visualization technique overlays a heatmap on the input image to show which regions most influenced a CNN's prediction?

Grad-CAM (Gradient-weighted Class Activation Mapping) (Correct answer)
t-SNE embedding
PCA projection
Batch normalization statistics

Correct answer: Grad-CAM (Gradient-weighted Class Activation Mapping)

Grad-CAM uses the gradients of the target class flowing into the final convolutional layer to produce a localization heatmap.

Question 4: What is the role of the activation function applied after a convolutional layer?

Normalizes the weights of the filter
Reduces the spatial dimensions of the feature map
Converts the feature map to a scalar
Introduces non-linearity so the network can learn complex feature hierarchies (Correct answer)

Correct answer: Introduces non-linearity so the network can learn complex feature hierarchies

Without non-linear activations, stacking convolutional layers would collapse to a single linear operation, losing representational power.

Question 5: In EfficientNet, what is 'compound scaling' and what does it scale jointly?

Scaling only depth to add more residual blocks uniformly
Scaling the learning rate and batch size together during training
Jointly scaling network width, depth, and input resolution using a fixed ratio (Correct answer)
Alternating between scaling width and depth in even and odd layers

Correct answer: Jointly scaling network width, depth, and input resolution using a fixed ratio

EfficientNet's compound scaling coefficient φ uniformly scales width (channels), depth (layers), and resolution (input size) with empirically derived ratios.

Question 6: What is 'Nesterov Accelerated Gradient' (NAG) and how does it differ from standard momentum SGD?

NAG applies momentum after the gradient step, while standard momentum applies it before
NAG computes the gradient at the lookahead position (after the momentum step) rather than the current position (Correct answer)
NAG adapts momentum based on curvature information
NAG uses second-order gradient information unlike standard momentum

Correct answer: NAG computes the gradient at the lookahead position (after the momentum step) rather than the current position

NAG evaluates the gradient at the predicted future position (current weights plus momentum step) rather than the current position, providing a more informed and corrective update.

Question 7: In a transposed convolution (sometimes called deconvolution), what is the typical use case in CNNs?

Spatial upsampling to increase feature map resolution (Correct answer)
Applying convolution in the frequency domain
Reversing gradient flow during backpropagation
Reducing depth of feature maps

Correct answer: Spatial upsampling to increase feature map resolution

Transposed convolutions learn to upsample feature maps and are used in decoder networks, GANs, and semantic segmentation architectures.

Question 8: What does 'stride' control in the context of the backbone's impact on object detection?

The learning rate decay schedule
The number of output classes
The spatial resolution of the output feature map relative to the input (Correct answer)
The number of anchor boxes per location

Correct answer: The spatial resolution of the output feature map relative to the input

Stride determines how much the feature map is downsampled; a stride of 32 means each feature map cell corresponds to a 32×32 pixel region in the input.

Question 9: GoogLeNet includes auxiliary classifiers during training. What is their primary purpose?

To create an ensemble of predictions at inference
To increase the number of output classes
To combat vanishing gradients by injecting gradient signal at intermediate layers (Correct answer)
To regularize the network using a multi-task loss

Correct answer: To combat vanishing gradients by injecting gradient signal at intermediate layers

Auxiliary classifiers in GoogLeNet inject gradient signal into earlier layers during backpropagation, helping train the deep network by fighting vanishing gradients.

Question 10: Dialysate glucose must be heat sterilized at a low pH to

Increase generation of glucose degradation products
Prevent the dialysate from becoming cloudy
Decrease generation of glucose degradation products (Correct answer)
Prevent crystallization in the dialysate

Correct answer: Decrease generation of glucose degradation products

Dialysate glucose must be heat sterilized at a low pH to minimize the formation of glucose degradation products (GDPs). High temperatures combined with a neutral or alkaline pH during sterilization can cause glucose to break down into harmful GDPs. These GDPs are associated with inflammation and damage to the peritoneal membrane in peritoneal dialysis patients, so maintaining a low pH helps preserve the integrity of the dialysate.

Question 11: What is the purpose of random rotation augmentation in CNN training?

To normalize pixel values
To reduce image dimensions
To make the model invariant to the orientation of objects in the image (Correct answer)
To speed up training convergence

Correct answer: To make the model invariant to the orientation of objects in the image

Random rotation augmentation rotates training images by a random angle, teaching the CNN to recognize objects regardless of their orientation.

Question 12: What is 'inverted dropout' and why is it preferred in practice?

Randomly scaling activations up instead of zeroing them down
Using a dropout rate that increases with network depth
Scaling kept activations by 1/(1-p) during training so no scaling is needed at test time (Correct answer)
Applying dropout only to the input layer and inverting the mask at deeper layers

Correct answer: Scaling kept activations by 1/(1-p) during training so no scaling is needed at test time

Inverted dropout scales kept activations by 1/(1-p) during training, so the network's expected output magnitude is unchanged and no scaling adjustment is needed at inference time.

Question 13: Which term describes sharing the same filter weights across all spatial positions in a convolutional layer?

Translation invariance
Sparse connectivity
Weight sharing (parameter sharing) (Correct answer)
Equivariance

Correct answer: Weight sharing (parameter sharing)

Weight sharing means the same filter is applied at every position, dramatically reducing parameters compared to a fully connected layer.

Question 14: Which data augmentation technique mixes two training images and their labels by linear interpolation to improve CNN generalization?

AutoAugment
CutMix
CutOut
Mixup (Correct answer)

Correct answer: Mixup

Mixup creates virtual training samples by interpolating pixel values and labels of two random images, encouraging the model to learn more linear behavior between classes.

Question 15: What is the output depth (number of channels) when 128 filters are applied to a 64-channel input?

192
64
8192
128 (Correct answer)

Correct answer: 128

The output depth always equals the number of filters, regardless of the input channel count.

Question 16: Which pooling operation is most commonly used in CNNs and what does it compute?

Average pooling — computes the mean of values in the pooling window
L2 pooling — computes the Euclidean norm of values in the window
Stochastic pooling — randomly samples a value weighted by activation magnitude
Max pooling — selects the maximum value in the pooling window (Correct answer)

Correct answer: Max pooling — selects the maximum value in the pooling window

Max pooling takes the largest activation within each region, retaining the strongest feature response and providing a degree of translation invariance.

Question 17: Which layer type in a CNN is responsible for introducing non-linearity?

Fully connected layer
Activation layer (Correct answer)
Normalization layer
Pooling layer

Correct answer: Activation layer

Activation layers (e.g., applying ReLU) introduce non-linearity after convolution, enabling the network to learn complex patterns.

Question 18: What is the primary effect of average pooling compared to max pooling in CNNs?

Increases spatial resolution of the feature map
Smooths feature maps by computing the mean of values in each region (Correct answer)
Produces sharper feature maps by emphasizing dominant activations
Eliminates negative activations from the feature map

Correct answer: Smooths feature maps by computing the mean of values in each region

Average pooling computes the mean of all values in a pooling window, resulting in smoother feature maps compared to max pooling.

Question 19: How does a transposed convolution (sometimes called deconvolution) differ from a standard convolution in terms of spatial output?

It operates along the channel axis rather than the spatial axes
It increases spatial dimensions, making it useful for upsampling (Correct answer)
It maintains identical spatial dimensions using same padding
It reduces spatial dimensions like a strided convolution

Correct answer: It increases spatial dimensions, making it useful for upsampling

Transposed convolution is the gradient operation of a forward convolution; it increases spatial size and is used in decoders, GANs, and segmentation networks.

Question 20: In VGGNet, what architectural choice was a key design principle?

Introducing residual skip connections between layers
Applying global average pooling before the classifier
Using a single large 11×11 kernel in the first layer
Stacking many small 3×3 convolutional filters instead of large kernels (Correct answer)

Correct answer: Stacking many small 3×3 convolutional filters instead of large kernels

VGGNet demonstrated that deep networks using only 3×3 convolutions can achieve strong performance, as two 3×3 layers have the same receptive field as one 5×5 layer with fewer parameters.

Question 21: Which augmentation strategy is specifically designed to force CNNs to use all object parts rather than the most discriminative region, by pasting patches between images?

CutMix (Correct answer)
Mixup
Mosaic
RandomCrop

Correct answer: CutMix

CutMix replaces a rectangular region of one training image with a patch from another image and blends labels proportionally, forcing the model to use the whole image.

Question 22: What is the primary advantage of using depthwise convolutions over standard convolutions in terms of computational cost?

They require more memory but execute faster on GPU
They eliminate the need for activation functions
They allow arbitrarily large kernel sizes without extra cost
They reduce multiply-add operations by separating spatial and channel-wise filtering (Correct answer)

Correct answer: They reduce multiply-add operations by separating spatial and channel-wise filtering

Depthwise convolutions apply one filter per input channel independently, reducing the number of multiply-add operations by a factor roughly equal to the number of output channels.

Question 23: Which of the following classic CNN architectures was the first to successfully demonstrate that significantly deeper networks could be trained by using 'skip connections' or 'residual blocks' to address the vanishing gradient problem?

VGG-16
ResNet (Correct answer)
LeNet-5
AlexNet

Correct answer: ResNet

ResNet (Residual Network) introduced the concept of residual blocks with skip connections. These connections create a shortcut path for the gradient to flow through, which helps mitigate the vanishing gradient problem that plagues very deep networks. This innovation allowed for the successful training of networks that were substantially deeper (e.g., 50, 101, or 152 layers) than previous architectures like AlexNet or VGG.

Question 24: Which scenario demonstrates 'double descent' behavior in CNN training?

Validation accuracy drops twice during learning rate warm-up
Loss oscillates when two different optimizers are combined
Training loss increases then decreases as epochs progress
Validation loss decreases, then increases, then decreases again as model size grows beyond the interpolation threshold (Correct answer)

Correct answer: Validation loss decreases, then increases, then decreases again as model size grows beyond the interpolation threshold

Double descent describes how test error can drop a second time as model size increases past the point of interpolation, counter to the classical bias-variance tradeoff.

Question 25: During backpropagation through a convolutional layer, what operation is performed to compute gradients with respect to the input?

Full convolution of the upstream gradient with the flipped filter (Correct answer)
Pooling of upstream gradients
Transposing the weight matrix and multiplying
Element-wise multiplication of the filter and upstream gradient

Correct answer: Full convolution of the upstream gradient with the flipped filter

The gradient with respect to the input is computed via convolution of the upstream gradient with the filter rotated 180°, which is equivalent to a full convolution.

Question 26: What is the receptive field of a neuron in the output of a convolutional layer?

The region of the input that influences that neuron's activation (Correct answer)
The number of input channels
The output feature map dimensions
The size of the filter

Correct answer: The region of the input that influences that neuron's activation

The receptive field is the spatial region in the input (or previous feature map) that a given output neuron 'sees' and is influenced by.

Question 27: What is 'catastrophic forgetting' in CNN training and which approach helps mitigate it?

The tendency of a neural network to forget previously learned tasks when trained on new data; mitigated by Elastic Weight Consolidation (EWC) (Correct answer)
Forgetting to normalize inputs; mitigated by batch normalization
Loss of training data due to hardware failure; mitigated by checkpointing
Gradient explosion causing weight overflow; mitigated by clipping

Correct answer: The tendency of a neural network to forget previously learned tasks when trained on new data; mitigated by Elastic Weight Consolidation (EWC)

Catastrophic forgetting occurs in continual learning when training on new tasks overwrites weights critical for old tasks; EWC adds a regularization term that penalizes changes to important weights.

Question 28: Which of the following best describes the role of the convolutional layer in a CNN?

To reduce the number of parameters in the network through downsampling.
To flatten the multi-dimensional feature maps into a one-dimensional vector.
To classify the input by connecting every neuron from the previous layer.
To apply a set of learnable filters to the input data to create feature maps. (Correct answer)

Correct answer: To apply a set of learnable filters to the input data to create feature maps.

The convolutional layer is the core building block of a CNN. Its primary function is to apply a series of learnable filters (or kernels) that slide over the input image to detect specific features like edges, corners, and textures, resulting in feature maps.

Question 29: A convolutional layer has 64 filters of size 3×3 applied to a 3-channel input. How many learnable parameters does it have (excluding bias)?

1728 (Correct answer)
576
9216
192

Correct answer: 1728

Parameters = filters × kernel_h × kernel_w × input_channels = 64 × 3 × 3 × 3 = 1728.

Question 30: What is the spatial dimension relationship between an input feature map and output feature map when using a 3×3 filter with padding='same' and stride=1?

Output size depends only on the number of filters
Output is the same spatial size as the input (Correct answer)
Output is larger by 2 in each dimension
Output is smaller by 2 in each dimension

Correct answer: Output is the same spatial size as the input

With 'same' padding and stride=1, zero-padding is added so the output spatial dimensions exactly match the input dimensions.

Question 31: What is the output size formula for a convolutional layer given input W, filter F, padding P, and stride S?

(W + F - 2P) / S - 1
(W - F + 2P) * S
(W * F) / (P + S)
(W + 2P - F) / S + 1 (Correct answer)

Correct answer: (W + 2P - F) / S + 1

The spatial output size is computed as (W + 2P − F) / S + 1, which accounts for padding, filter size, and the step size.

Question 32: The process in a convolutional layer where a filter slides over the input data, computes element-wise products, and sums them up to create a single value in the output is known as what?

Flattening
Activation
Convolution (Correct answer)
Max Pooling

Correct answer: Convolution

This describes the fundamental convolution operation. The filter (or kernel) moves across the input, and at each position, the dot product between the filter's weights and the corresponding input values is calculated. This result forms one element of the output feature map.

Question 33: How does dropout act as a form of regularization?

By increasing the effective learning rate to escape sharp minima
By forcing the network to learn robust features not dependent on specific co-adaptations (Correct answer)
By reducing the number of parameters in the network permanently
By penalizing large weight values via an L2 term in the loss function

Correct answer: By forcing the network to learn robust features not dependent on specific co-adaptations

Dropout prevents co-adaptation of neurons by randomly removing them, forcing each neuron to learn useful features independently, which improves generalization.

Question 34: Which CNN training technique uses multiple GPUs where each device holds a copy of the model and processes a different mini-batch?

Pipeline parallelism
Tensor parallelism
Model parallelism
Data parallelism (Correct answer)

Correct answer: Data parallelism

Data parallelism replicates the model on each GPU and splits mini-batches across devices, synchronizing gradients after each forward-backward pass.

Question 35: What is dilated (atrous) convolution, and what advantage does it provide?

A convolution that uses fractional strides
A convolution applied only to dilated images
A convolution with gaps between filter elements, expanding receptive field without adding parameters (Correct answer)
A convolution with larger filters and more parameters

Correct answer: A convolution with gaps between filter elements, expanding receptive field without adding parameters

Dilation inserts zeros between filter elements (dilation rate > 1), covering a larger area of the input with the same number of parameters.

Question 36: A data scientist is training a CNN and wants to implement a learning rate schedule that starts with a relatively high learning rate and then smoothly decreases it following the shape of a cosine curve, potentially with periodic restarts. What is this scheduling strategy called?

Exponential Decay
Step Decay
Time-Based Decay
Cosine Annealing (Correct answer)

Correct answer: Cosine Annealing

Cosine Annealing is a learning rate schedule where the learning rate is adjusted according to the cosine function. It starts with a higher value and smoothly 'anneals' or decreases to a minimum value. [26] This strategy can be very effective, especially when used with 'warm restarts,' where the learning rate is periodically reset to its initial high value, which can help the model escape poor local minima. [26]

Question 37: How does a 1×1 convolution (pointwise convolution) affect a feature map?

It applies spatial filtering across a local region
It increases spatial resolution
It changes the number of channels without altering spatial dimensions (Correct answer)
It acts as a pooling operation

Correct answer: It changes the number of channels without altering spatial dimensions

A 1×1 convolution performs a linear combination across channels at each spatial position, enabling channel dimensionality reduction or expansion.

Question 38: What does the CutOut augmentation technique do?

Removes the image background
Crops the image into multiple patches
Applies random color cutoffs
Randomly masks out a square region of the input during training (Correct answer)

Correct answer: Randomly masks out a square region of the input during training

CutOut randomly zeroes out a square patch in the training image, forcing the network to use context from the entire image rather than relying on a single discriminative region.

Question 39: In deep CNN training, what is 'residual learning' (as in ResNets) and how does it aid optimization?

Using residual blocks to reduce the number of parameters
Learning the difference from a target distribution using KL divergence
Subtracting mean activations to normalize feature distributions
Learning residual functions with reference to layer inputs via skip connections, easing gradient flow (Correct answer)

Correct answer: Learning residual functions with reference to layer inputs via skip connections, easing gradient flow

Residual learning reformulates layers to learn residual functions F(x) added back to the input x, making identity mappings easy and allowing gradient to flow directly through skip connections to earlier layers.

Question 40: What pooling strategy does GoogLeNet use at the end of the network instead of fully connected layers?

Fractional max pooling
Max pooling
Average pooling (Correct answer)
Adaptive average pooling

Correct answer: Average pooling

GoogLeNet uses global average pooling before the final classifier, drastically reducing parameters compared to flattening into large fully connected layers.

Question 41: What is the 'bias-variance tradeoff' as it applies to CNN design?

Larger datasets always reduce both bias and variance equally
Regularization increases both bias and variance
Increasing model capacity reduces bias but can increase variance (overfitting) (Correct answer)
Adding more layers always reduces both bias and variance

Correct answer: Increasing model capacity reduces bias but can increase variance (overfitting)

Larger CNN capacity reduces training error (bias) but increases sensitivity to training data fluctuations (variance), risking overfitting.

Question 42: A key innovation of the GoogLeNet (Inception v1) architecture was the 'Inception module'. What is the primary purpose of this module?

To drastically increase the network depth by using residual connections.
To perform convolutions at multiple scales (1x1, 3x3, 5x5) in parallel and concatenate the results. (Correct answer)
To simplify the architecture by using only 3x3 convolutions.
To replace all fully connected layers with global average pooling to reduce parameters.

Correct answer: To perform convolutions at multiple scales (1x1, 3x3, 5x5) in parallel and concatenate the results.

The core idea of the Inception module in GoogLeNet is to allow the network to capture features at multiple scales simultaneously. It achieves this by performing 1x1, 3x3, and 5x5 convolutions, along with a max pooling operation, in parallel within the same module. The outputs are then concatenated, creating a rich, multi-scale feature representation.

Question 43: What is the role of skip connections in the U-Net architecture?

They skip residual computations to speed up training
They skip low-performing layers during inference
They pass high-resolution feature maps from the encoder to the decoder to recover spatial detail lost during downsampling (Correct answer)
They connect non-adjacent layers for gradient flow only

Correct answer: They pass high-resolution feature maps from the encoder to the decoder to recover spatial detail lost during downsampling

Skip connections concatenate encoder feature maps with decoder feature maps at matching resolutions, allowing the decoder to recover fine spatial details that are lost during pooling.

Question 44: A data scientist is designing a CNN to classify high-resolution medical images. They are concerned about losing important information at the borders of the images during the convolution operations. Which technique should they employ to mitigate this issue?

Increasing the stride of the convolution.
Adding a padding layer around the input images. (Correct answer)
Using a larger pooling window.
Implementing a dropout layer after the convolutional layer.

Correct answer: Adding a padding layer around the input images.

Padding involves adding extra pixels (usually zeros) around the border of an input image. This technique ensures that the filter can process the pixels at the edges of the image more effectively, preventing the spatial dimensions from shrinking with each convolution and preserving information at the borders.

Question 45: What is the primary purpose of data augmentation in CNN training?

To artificially expand the training dataset and improve generalization (Correct answer)
To increase model size
To speed up inference
To reduce the number of layers

Correct answer: To artificially expand the training dataset and improve generalization

Data augmentation artificially expands the training dataset by applying transformations, helping CNNs generalize better to unseen data.

Question 46: During CNN training, which regularization method randomly drops entire feature maps rather than individual neurons?

SpatialDropout
DropBlock (Correct answer)
Weight decay
Dropout

Correct answer: DropBlock

DropBlock drops contiguous regions of feature maps, which is more effective for CNNs than standard dropout because adjacent units in feature maps tend to be correlated.

Question 47: What does 'gradient accumulation' allow during CNN training with limited GPU memory?

Caching gradients to speed up the backward pass
Combining gradients from multiple models for ensemble training
Using a larger effective batch size by accumulating gradients over multiple mini-batches before updating (Correct answer)
Storing gradients across epochs to avoid recomputation

Correct answer: Using a larger effective batch size by accumulating gradients over multiple mini-batches before updating

Gradient accumulation sums gradients over several small mini-batches before performing a weight update, simulating a larger batch size without requiring additional GPU memory.

Question 48: How does GoogLeNet reduce computational cost before applying 3x3 and 5x5 convolutions in the Inception module?

By applying 1x1 convolutions as bottlenecks (Correct answer)
By reducing the input image resolution
By using strided convolutions
By using average pooling before convolution

Correct answer: By applying 1x1 convolutions as bottlenecks

1x1 convolutions act as bottlenecks that reduce the number of input channels before more expensive 3x3 and 5x5 convolutions are applied.

Question 49: During CNN training, what is the purpose of the 'learning rate warmup' phase?

To reduce batch size at the start of training
To freeze early layers while training later layers first
To gradually increase the learning rate from near-zero before using the main schedule (Correct answer)
To start with a large learning rate for fast initial progress

Correct answer: To gradually increase the learning rate from near-zero before using the main schedule

Warmup slowly increases the learning rate from a very small value, stabilizing early training when weights and gradients are poorly scaled before the main learning rate schedule begins.

Question 50: What role does global average pooling (GAP) play as a regularizer compared to using fully connected layers before the output?

GAP eliminates the need for large fully connected layers, drastically reducing overfitting-prone parameters (Correct answer)
GAP applies L1 regularization to the final feature map
GAP adds more parameters to the model
GAP increases gradient magnitude during backpropagation

Correct answer: GAP eliminates the need for large fully connected layers, drastically reducing overfitting-prone parameters

GAP replaces large fully connected layers by averaging each feature map to a single value, dramatically reducing parameter count and overfitting risk.

Question 51: What is the effect of applying batch normalization after a convolutional layer before the activation function?

Replaces the bias term with a learned mean
Doubles the number of parameters in the layer
Normalizes activations to have zero mean and unit variance, stabilizing training (Correct answer)
Removes the need for filters entirely

Correct answer: Normalizes activations to have zero mean and unit variance, stabilizing training

Batch normalization standardizes pre-activation values across the mini-batch, reducing internal covariate shift and enabling higher learning rates.

Question 52: What is the primary purpose of padding in a convolutional layer?

To normalize pixel values before convolution
To reduce the number of filters
To increase the depth of the feature maps
To preserve spatial dimensions of the input (Correct answer)

Correct answer: To preserve spatial dimensions of the input

Padding adds border pixels (usually zeros) around the input so the output feature map retains the same spatial size as the input.

Question 53: A CNN uses L2 regularization with coefficient λ. What happens to the gradient update rule for a weight w?

The gradient is divided by λ at each step
The weight is clipped to [−λ, λ] after each update
An additive penalty λ is added to the loss, and the gradient gains a +2λw term (Correct answer)
A term −2λw is subtracted from the weight before the gradient step

Correct answer: An additive penalty λ is added to the loss, and the gradient gains a +2λw term

L2 regularization adds λ‖w‖² to the loss, so its gradient contribution is +2λw, which is added to the usual gradient, effectively shrinking weights.

Question 54: A convolutional layer in a CNN has an input of size 64x64x16. It uses 32 filters, each of size 3x3, a stride of 1, and 'same' padding. What is the dimensionality of the output feature map?

62x62x32
64x64x32 (Correct answer)
32x32x32
64x64x16

Correct answer: 64x64x32

The output dimensions are calculated as follows: Output Width = (Input Width - Filter Width + 2 * Padding) / Stride + 1. With 'same' padding, the output height and width are preserved to be the same as the input (64x64). The depth of the output is determined by the number of filters used, which is 32. Therefore, the output volume is 64x64x32.

Question 55: When stride > 1 is used in a convolutional layer, what is the primary effect?

Wider receptive field per filter step
Spatial downsampling of the feature map (Correct answer)
Spatial upsampling of the feature map
Increased number of parameters

Correct answer: Spatial downsampling of the feature map

A larger stride causes the filter to skip positions, reducing the output spatial dimensions (downsampling).

Question 56: What is the role of anchor boxes in SSD and Faster R-CNN?

They replace the need for a backbone network
They determine the learning rate for each layer
They define the output image resolution
They serve as reference boxes of predefined scales and aspect ratios for regression (Correct answer)

Correct answer: They serve as reference boxes of predefined scales and aspect ratios for regression

Anchor boxes provide predefined reference shapes that the network adjusts via regression to match actual object locations and sizes.

Question 57: Which technique adds Gaussian noise to the gradients during CNN training to help escape sharp local minima?

Weight decay
Batch normalization
Gradient clipping
Stochastic gradient noise (Correct answer)

Correct answer: Stochastic gradient noise

Stochastic gradient noise injects Gaussian noise into gradients during optimization, helping the model escape sharp local minima and find flatter, more generalizable optima.

Question 58: In the Feature Pyramid Network (FPN), how are features from different levels combined?

By average pooling across levels
By top-down pathway with lateral connections (Correct answer)
By concatenation only
By a transformer attention mechanism

Correct answer: By top-down pathway with lateral connections

FPN uses a top-down pathway that upsamples high-level features and adds them to lower-level features via lateral connections.

Question 59: What distinguishes instance segmentation from semantic segmentation in the context of object detection?

Instance segmentation only works on single-object images
Instance segmentation distinguishes between separate instances of the same class (Correct answer)
Semantic segmentation provides bounding boxes while instance segmentation does not
Instance segmentation uses only fully connected layers

Correct answer: Instance segmentation distinguishes between separate instances of the same class

Instance segmentation assigns a unique mask to each individual object instance, while semantic segmentation labels each pixel with a class without distinguishing instances.

Question 60: In CNN training, what is 'batch size' effect on generalization — specifically why do smaller batches often generalize better?

Smaller batches introduce more gradient noise acting as implicit regularization (Correct answer)
Smaller batches converge faster to the global minimum
Smaller batches compute more accurate gradients
Smaller batches reduce overfitting by using less data per step

Correct answer: Smaller batches introduce more gradient noise acting as implicit regularization

Small batch SGD introduces higher gradient variance (noise), which acts as implicit regularization and tends to find flatter minima that generalize better than sharp minima found by large batches.

Question 61: What is the key innovation in MobileNetV2's inverted residual block compared to standard residual blocks?

It uses dilated convolutions to increase receptive field without more parameters
It applies group convolutions to split channels into independent subgroups
It replaces ReLU with sigmoid to prevent dead neurons
It expands channels before the depthwise convolution, then compresses with a linear bottleneck (Correct answer)

Correct answer: It expands channels before the depthwise convolution, then compresses with a linear bottleneck

MobileNetV2 inverts the bottleneck by expanding channels first, applying depthwise conv, then projecting back to a low-dimensional space with a linear activation.

Question 62: What distinguishes a 'same' padding strategy from a 'valid' padding strategy in convolutional layers?

Same padding pads the input so output spatial size equals input size; valid padding applies no padding, shrinking output (Correct answer)
Same padding is only used in pooling; valid padding applies to convolutions
Same padding doubles the input size; valid padding halves it
Same padding uses larger filters; valid padding uses 1×1 filters

Correct answer: Same padding pads the input so output spatial size equals input size; valid padding applies no padding, shrinking output

With 'same' padding, zeros are added so the output matches the input's spatial dimensions; with 'valid', no padding is added and the output shrinks based on filter size.

Question 63: In a fully convolutional network (FCN), what replaces the fully connected layers to enable pixel-wise predictions?

Global max pooling across spatial dimensions
Additional pooling layers to collapse the feature map
Transposed convolutions (or bilinear upsampling) to recover spatial resolution (Correct answer)
Dropout layers applied at every spatial location

Correct answer: Transposed convolutions (or bilinear upsampling) to recover spatial resolution

FCNs replace dense layers with convolutional ones and use transposed convolutions to upsample feature maps back to the input resolution for dense prediction.

Question 64: What does a feature map (activation map) represent in a CNN layer?

The raw pixel values of the input image
The output of applying a learned filter across the input (Correct answer)
The gradient of the loss with respect to weights
The pooled output after max pooling only

Correct answer: The output of applying a learned filter across the input

A feature map is produced by convolving a learned filter over the input, showing where and how strongly that filter's pattern is detected.

Question 65: A practitioner reduces the number of filters in each CNN convolutional layer by half. How does this combat overfitting?

It increases the effective learning rate
It applies L1 penalty to activations
It reduces total model parameters, lowering capacity to memorize noise (Correct answer)
It increases training data diversity

Correct answer: It reduces total model parameters, lowering capacity to memorize noise

Fewer filters means fewer learnable parameters, directly reducing the model's capacity to overfit.

Question 66: What is 'cosine annealing' in CNN learning rate scheduling?

Oscillating the learning rate between two values sinusoidally
Linearly decreasing the learning rate each epoch
Multiplying the learning rate by a constant factor at fixed intervals
Decreasing the learning rate following a cosine curve from maximum to minimum (Correct answer)

Correct answer: Decreasing the learning rate following a cosine curve from maximum to minimum

Cosine annealing smoothly reduces the learning rate following half a cosine cycle from an initial maximum to a minimum value, enabling gradual and smooth convergence.

Question 67: What is the 'dead ReLU' problem during CNN training?

ReLU causes exploding gradients in deep networks
ReLU causes the loss to become non-differentiable
Neurons with negative pre-activation permanently output zero and stop learning (Correct answer)
ReLU activation leads to sparse but unreliable features

Correct answer: Neurons with negative pre-activation permanently output zero and stop learning

Dead ReLUs occur when a neuron's input is always negative, causing it to output zero and receive zero gradient, effectively removing it from learning permanently.

Question 68: In a typical Convolutional Neural Network (CNN) architecture, what is the primary function of the pooling layer?

To perform classification based on extracted features.
To introduce non-linearity into the model.
To reduce the spatial dimensions of the feature maps. (Correct answer)
To extract features like edges and textures from the input image.

Correct answer: To reduce the spatial dimensions of the feature maps.

The pooling layer, also known as a downsampling layer, is primarily used to reduce the spatial dimensions (width and height) of the input feature maps. This process helps to decrease the computational complexity, control overfitting, and create an invariance to small translations in the input image.

Question 69: What is the primary role of the Region Proposal Network (RPN) in Faster R-CNN?

Perform non-maximum suppression
Propose candidate object bounding boxes (Correct answer)
Extract feature maps from the backbone
Generate class probability scores

Correct answer: Propose candidate object bounding boxes

The RPN slides over the feature map and proposes candidate bounding boxes (anchors) that may contain objects.

Question 70: In the context of a convolutional operation, what does the 'stride' parameter define?

The number of pixels by which the filter slides over the input at each step. (Correct answer)
The size of the pooling window used for downsampling.
The amount of zero-padding added to the borders of the input.
The number of filters applied to the input volume.

Correct answer: The number of pixels by which the filter slides over the input at each step.

The stride defines the step size the convolutional filter moves across the input image. A stride of 1 means the filter moves one pixel at a time. A larger stride (e.g., 2) means the filter jumps 2 pixels at each step, resulting in a smaller output feature map and reduced computation.

Question 71: What problem do residual connections in ResNet primarily solve?

Slow inference speed during deployment
Excessive memory usage from large feature maps
Vanishing gradients that hinder training of very deep networks (Correct answer)
Overfitting on small datasets

Correct answer: Vanishing gradients that hinder training of very deep networks

Residual (skip) connections allow gradients to flow directly through the network, mitigating vanishing gradient problems that arise in very deep architectures.

Question 72: What is the receptive field of a neuron in a CNN?

The output size of the layer it belongs to
The region of the input image that influences that neuron's activation (Correct answer)
The number of channels it processes
The set of weights in its corresponding filter

Correct answer: The region of the input image that influences that neuron's activation

The receptive field is the spatial extent of the original input that a given neuron 'sees', which grows with depth due to successive convolutions.

Question 73: What is the primary goal of using regularization techniques like Dropout or L2 regularization when training a Convolutional Neural Network?

To automatically determine the optimal number of convolutional layers.
To increase the model's performance on the training dataset.
To reduce the model's complexity and improve its ability to generalize to unseen data. (Correct answer)
To increase the speed of convergence during training.

Correct answer: To reduce the model's complexity and improve its ability to generalize to unseen data.

The fundamental purpose of regularization is to prevent overfitting. [2, 12] Overfitting occurs when a model learns the training data too well and fails to generalize to new, unseen data. [24, 27] Regularization techniques introduce constraints or penalties on the model's parameters (like L2 regularization) or its structure during training (like Dropout) to reduce its complexity and force it to learn more robust, generalizable patterns. [6, 11]

Question 74: What is the effect of applying Gaussian blur augmentation to CNN training images?

It sharpens edges in the image
It simulates out-of-focus conditions, making the model robust to image blurriness (Correct answer)
It increases image contrast
It converts the image to grayscale

Correct answer: It simulates out-of-focus conditions, making the model robust to image blurriness

Gaussian blur augmentation convolves the image with a Gaussian kernel, simulating defocus or motion blur and helping CNNs generalize to lower-quality input images.

Question 75: Towards the end of a CNN architecture, after the convolutional and pooling layers have extracted features, which layer is typically responsible for taking these high-level features and performing the final classification task?

Another convolutional layer
A fully connected (dense) layer (Correct answer)
A max-pooling layer
An activation layer like ReLU

Correct answer: A fully connected (dense) layer

The fully connected (or dense) layer takes the high-level features from the preceding layers (which are often flattened into a 1D vector) and performs the final classification. Each neuron in a fully connected layer is connected to all neurons in the previous layer, allowing it to learn non-linear combinations of these features to make a prediction.

Question 76: Why is batch normalization particularly important in semantic segmentation CNN training?

It augments the training data automatically
It increases the spatial resolution of feature maps
It reduces the number of output classes
It normalizes feature map activations across the batch to stabilize training and reduce sensitivity to weight initialization (Correct answer)

Correct answer: It normalizes feature map activations across the batch to stabilize training and reduce sensitivity to weight initialization

Batch normalization normalizes activations within each mini-batch, reducing internal covariate shift and enabling higher learning rates, which is critical for the deep encoder-decoder architectures common in segmentation.

Question 77: Which technique involves training the CNN with the true label replaced by a soft distribution over all classes to reduce overconfidence?

Temperature scaling
Knowledge distillation
Mixup
Label smoothing (Correct answer)

Correct answer: Label smoothing

Label smoothing replaces hard one-hot targets with soft targets (e.g., 0.9 for correct class, 0.1/K for others) to prevent over-confident predictions.

Question 78: What is Panoptic Segmentation?

Segmentation applied to panoramic images only
A multi-view segmentation approach
Segmentation applied exclusively to medical images
A unified task that combines semantic segmentation (stuff) and instance segmentation (things) into a single cohesive map (Correct answer)

Correct answer: A unified task that combines semantic segmentation (stuff) and instance segmentation (things) into a single cohesive map

Panoptic segmentation assigns each pixel both a semantic class label and an instance ID where applicable, unifying the 'stuff' categories of semantic segmentation with the 'things' categories of instance segmentation.

Question 79: In AlexNet, what was the purpose of Local Response Normalization (LRN)?

To normalize activations across neighboring feature maps for lateral inhibition (Correct answer)
To replace dropout during training
To scale gradients during backpropagation
To reduce the spatial size of feature maps

Correct answer: To normalize activations across neighboring feature maps for lateral inhibition

LRN in AlexNet suppresses activations that are weak relative to neighbors across channels, mimicking lateral inhibition seen in biological neurons.

Question 80: What filter size does VGGNet use exclusively in its convolutional layers?

7x7
1x1
5x5
3x3 (Correct answer)

Correct answer: 3x3

VGGNet uses only 3x3 convolutional filters, showing that stacking small filters achieves the same receptive field as larger filters with fewer parameters.

Question 81: In CNN training, what does 'warm restarts' refer to in the context of learning rate schedules?

Gradually increasing batch size
Restarting training from scratch periodically
Periodically resetting the learning rate to a high value then annealing it (Correct answer)
Warming up GPU memory before training

Correct answer: Periodically resetting the learning rate to a high value then annealing it

Warm restarts (SGDR) periodically resets the learning rate to a maximum value and then follows a cosine annealing schedule, helping the model explore different loss landscape regions.

Question 82: Which component of Faster R-CNN is shared between the RPN and the detection head?

The bounding box regression layer
The convolutional feature extractor (backbone) (Correct answer)
The RoI pooling layer
The classification layer

Correct answer: The convolutional feature extractor (backbone)

Both the RPN and the detection head use the same convolutional backbone features, making computation efficient.

Question 83: Melissa Cheng, a 62-year-old woman receiving hemodialysis, complains of constant scratching. <br> <br> Question: If replacing Ms Walker's dialyzers and modifying her Kt/V do not relieve her itching, the most recommended step is:

Tacrolimus ointment
Gabapentin
Moisturizers/oil bath (Correct answer)
UBV phototherapy

Correct answer: Moisturizers/oil bath

Uremic pruritus, or itching, is a common and distressing symptom in hemodialysis patients. After optimizing dialysis adequacy (Kt/V) and ruling out other causes, the initial and most conservative approach to managing itching is often topical treatment. Moisturizers and oil baths help to hydrate the skin, reduce dryness, and alleviate irritation, which can significantly reduce the severity of pruritus.

Question 84: What is the effect of using a very large learning rate when training a CNN with batch normalization?

It causes gradient values to be clipped by BN, preventing learning
It has no effect because BN removes the dependence on learning rate scale
It immediately causes NaN losses due to BN's variance computation
It can be tolerated better than without BN because BN stabilizes the distribution of layer inputs (Correct answer)

Correct answer: It can be tolerated better than without BN because BN stabilizes the distribution of layer inputs

Batch Normalization reduces sensitivity to learning rate by normalizing activations, allowing higher rates that would cause instability in unnormalized networks.

Question 85: Which of the following activation functions would be most appropriate for the output layer of a CNN designed for a multi-class classification problem (e.g., classifying images into 10 different categories)?

ReLU
Softmax (Correct answer)
Sigmoid
Tanh

Correct answer: Softmax

The Softmax function is ideal for multi-class classification because it converts a vector of raw output scores (logits) into a probability distribution over the classes. Each output value is between 0 and 1, and the sum of all output values equals 1, representing the model's confidence for each class.

Question 86: What is global average pooling (GAP) and why is it used before the final classifier in modern CNNs?

It pools across the channel dimension to reduce depth
It applies max pooling with a kernel equal to the full feature map size
It averages across spatial dimensions to produce a single value per feature map, replacing dense layers (Correct answer)
It normalizes activations globally to prevent covariate shift

Correct answer: It averages across spatial dimensions to produce a single value per feature map, replacing dense layers

GAP collapses each feature map to a scalar by averaging all spatial values, drastically reducing parameters and providing some translation invariance.

Question 87: Which classic CNN was specifically designed to fit on embedded systems with a model size under 0.5MB using 'fire modules'?

EfficientNet-B0
MobileNet
SqueezeNet (Correct answer)
ShuffleNet

Correct answer: SqueezeNet

SqueezeNet (2016) used 'fire modules' with squeeze and expand layers to achieve AlexNet-level accuracy at less than 0.5MB, targeting embedded deployment.

Question 88: Which innovation did Mask R-CNN add to Faster R-CNN?

A parallel branch for instance segmentation masks (Correct answer)
Depthwise separable convolutions in the backbone
Focal loss for class imbalance
Anchor-free bounding box prediction

Correct answer: A parallel branch for instance segmentation masks

Mask R-CNN adds a small FCN branch that predicts a binary segmentation mask for each detected object instance in parallel with the class and box heads.

Question 89: What does the term 'feature map' refer to in the context of a convolutional layer output?

The gradient map used during backpropagation
A 2D grid of activations produced by applying one filter to the input (Correct answer)
The flattened output before the fully connected layer
The learned weight matrix of a filter

Correct answer: A 2D grid of activations produced by applying one filter to the input

A feature map (or activation map) is the 2D spatial output produced by convolving one filter across the entire input.

Question 90: When both batch normalization and dropout are used in the same network, what ordering is generally recommended?

Conv → Dropout → Batch Norm → ReLU
Conv → Batch Norm → ReLU → Dropout (in fully connected layers) (Correct answer)
Batch Norm → Conv → Dropout → ReLU
Conv → ReLU → Dropout → Batch Norm

Correct answer: Conv → Batch Norm → ReLU → Dropout (in fully connected layers)

The standard practice is Conv → BN → ReLU for convolutional blocks, with dropout applied in fully connected layers after activation, as applying dropout before BN can disrupt the normalization statistics.

Question 91: Which Python library function is commonly used to extract intermediate layer outputs for feature map visualization in Keras?

layer.get_weights()
model.predict()
Model(inputs, intermediate_layer.output) (Correct answer)
keras.backend.function()

Correct answer: Model(inputs, intermediate_layer.output)

Creating a new Keras Model with the original input and an intermediate layer's output allows you to forward-pass an image and retrieve that layer's activations.

Question 92: Which of the following best describes the primary purpose of a 1x1 convolution operation in a CNN architecture?

To act as a channel-wise, fully connected layer for dimensionality reduction or expansion. (Correct answer)
To detect complex spatial features like edges and corners.
To significantly increase the receptive field of the network.
To perform spatial down-sampling similar to a pooling layer.

Correct answer: To act as a channel-wise, fully connected layer for dimensionality reduction or expansion.

A 1x1 convolution operates across all channels at a single pixel location. This allows it to learn combinations of features across channels, effectively acting like a fully connected layer for the channel dimension. It is widely used for dimensionality reduction (by using fewer filters than input channels) or expansion, which helps in building more efficient architectures like Inception networks.

Question 93: In the context of CNNs, what does 'weight sharing' mean and why is it important?

The same filter weights are applied at every spatial location, drastically reducing parameters (Correct answer)
Multiple layers share the same weights to save memory
Convolutional and fully connected layers reuse the same weight matrix
Weights are shared between the encoder and decoder paths

Correct answer: The same filter weights are applied at every spatial location, drastically reducing parameters

Weight sharing means one learned filter slides across the entire input, so a layer with a 3×3 filter has only 9 weights per channel regardless of input size.

Question 94: When fine-tuning only the last few layers of a pretrained CNN, what is the main concern if you use a very high learning rate?

Underfitting the new dataset
Overfitting the training data
Catastrophic forgetting of pretrained features (Correct answer)
Increased inference time

Correct answer: Catastrophic forgetting of pretrained features

A high learning rate can overwrite the valuable pretrained weights through catastrophic forgetting, destroying learned representations.

Question 95: CutOut augmentation reduces CNN overfitting by which mechanism?

Adding Gaussian noise to pixel values
Blending two images together
Flipping the image horizontally
Randomly masking out rectangular regions of training images (Correct answer)

Correct answer: Randomly masking out rectangular regions of training images

CutOut randomly removes square patches from training images, forcing the CNN to use multiple regions rather than relying on any single discriminative area.

Question 96: Which statement about LIME (Local Interpretable Model-agnostic Explanations) in the context of CNNs is correct?

LIME perturbs superpixels of the input and fits a local linear model to explain predictions (Correct answer)
LIME requires access to the model's internal gradients
LIME directly visualizes convolutional feature maps
LIME only works with fully connected networks

Correct answer: LIME perturbs superpixels of the input and fits a local linear model to explain predictions

LIME explains CNN predictions by masking image superpixels, observing prediction changes, and fitting a simple interpretable model to approximate local decision boundaries.

Question 97: What is Test-Time Augmentation (TTA) used for in CNN inference?

Applying multiple augmentations to test images and averaging predictions to improve accuracy (Correct answer)
Speeding up inference time
Reducing model parameters at test time
Augmenting the test dataset permanently

Correct answer: Applying multiple augmentations to test images and averaging predictions to improve accuracy

TTA applies several augmented versions of each test image, collects the model's predictions for each, and averages them to produce a more robust final prediction.

Question 98: When freezing layers during transfer learning in Keras/TensorFlow, which attribute is set to False?

layer.requires_grad
layer.active
layer.frozen
layer.trainable (Correct answer)

Correct answer: layer.trainable

Setting layer.trainable = False in Keras prevents weight updates for that layer during training.

Question 99: What is the primary motivation for 'freezing' the early convolutional layers of a pre-trained CNN when applying transfer learning to a new task?

To ensure that the optimizer can apply a uniform learning rate across all layers of the network.
To significantly increase the training speed by reducing the number of backpropagation calculations.
To preserve the generic, low-level feature detectors (e.g., edges, colors, textures) that the model learned from the original large-scale dataset. (Correct answer)
To force the model to learn new low-level features that are highly specific to the new, smaller dataset.

Correct answer: To preserve the generic, low-level feature detectors (e.g., edges, colors, textures) that the model learned from the original large-scale dataset.

The initial layers of a CNN learn to detect general features like edges, corners, and color blobs, which are applicable to most computer vision tasks. Freezing these layers prevents their weights from being updated, thereby preserving this fundamental knowledge and preventing it from being corrupted by training on a potentially small or different dataset.

Question 100: Which operation sums element-wise products between a filter and a receptive field in a CNN?

Matrix multiplication
Cross-correlation (Correct answer)
Dot product of flattened tensors
Convolution without flipping

Correct answer: Cross-correlation

Cross-correlation (often loosely called convolution in deep learning) computes the sum of element-wise products without flipping the filter.

Question 101: Why does max pooling provide a degree of translation invariance in CNNs?

It selects the peak activation in a region regardless of its exact position within that window (Correct answer)
It rotates the feature map to align dominant orientations
It averages activations, smoothing out positional differences
It normalizes activations so that their spatial positions become irrelevant

Correct answer: It selects the peak activation in a region regardless of its exact position within that window

Max pooling discards the precise location of a feature within the pooling window, so small shifts in input position produce the same output, yielding local translation invariance.

Question 102: What happens to the gradient of the Sigmoid activation function when its input is very large (e.g., |x| > 5)?

The gradient becomes negative, reversing weight update direction
The gradient approaches zero, causing the vanishing gradient problem (Correct answer)
The gradient becomes very large, causing exploding gradients
The gradient remains constant at 0.25, the maximum for Sigmoid

Correct answer: The gradient approaches zero, causing the vanishing gradient problem

Sigmoid saturates near 0 and 1 for large |x|, making its derivative close to zero and causing gradients to vanish during backpropagation through multiple layers.

Question 103: What is the fundamental difference between semantic segmentation and instance segmentation?

Instance segmentation works only on videos
Semantic segmentation requires depth data
Semantic segmentation assigns the same label to all pixels of a class, while instance segmentation distinguishes individual object instances (Correct answer)
Semantic segmentation is faster

Correct answer: Semantic segmentation assigns the same label to all pixels of a class, while instance segmentation distinguishes individual object instances

In semantic segmentation all pixels of the same class share one label, whereas instance segmentation distinguishes separate instances so two cars get different masks.

Question 104: Which classic CNN architecture introduced Local Response Normalization (LRN)?

VGGNet
AlexNet (Correct answer)
GoogLeNet
LeNet-5

Correct answer: AlexNet

AlexNet introduced Local Response Normalization, which normalizes across adjacent feature maps to encourage competition between neurons — though later research found it rarely beneficial.

Question 105: Which property of pooling operations makes CNNs more robust to small translations in the input image?

Pooling normalizes pixel intensities across the input
Pooling amplifies high-frequency details in feature maps
Pooling introduces translational invariance by summarizing local regions (Correct answer)
Pooling increases the number of learnable parameters

Correct answer: Pooling introduces translational invariance by summarizing local regions

By taking the max or average over a local region, pooling makes the output insensitive to small shifts of features within that region.

Question 106: What is the purpose of 'maximally activating patches' in CNN visualization?

To find input image regions that produce the highest activation for a specific filter (Correct answer)
To prune filters with low average activation
To reduce the number of channels in a layer
To normalize feature maps across a batch

Correct answer: To find input image regions that produce the highest activation for a specific filter

Maximally activating patches are the input image crops that cause a particular filter or neuron to fire most strongly, revealing what it has learned to detect.

Question 107: In DenseNet, how does each layer receive its input?

From a learned weighted sum of all preceding layers
From alternating layers spaced two apart
Only from the immediately preceding layer
From all preceding layers via concatenation (Correct answer)

Correct answer: From all preceding layers via concatenation

DenseNet concatenates feature maps from all preceding layers as input to each layer, maximizing feature reuse and enabling gradient flow through the entire network.

Question 108: What is the DeepLab v3+ architecture's improvement over earlier DeepLab versions?

It removes all dilated convolutions
It uses recurrent layers instead of convolutions
It replaces the backbone with a transformer
It adds a decoder module that refines segmentation boundaries by combining low-level encoder features with the ASPP output (Correct answer)

Correct answer: It adds a decoder module that refines segmentation boundaries by combining low-level encoder features with the ASPP output

DeepLab v3+ introduces a simple but effective decoder that upsamples the ASPP output and concatenates it with low-level features from the encoder, sharpening object boundaries compared to simple bilinear upsampling.

Question 109: Guided Backpropagation improves standard backpropagation visualization by doing what?

Replacing all activations with their absolute values
Using second-order gradients (Hessians) instead of first-order
Zeroing out negative gradients at ReLU gates during the backward pass (Correct answer)
Backpropagating only through the final fully connected layer

Correct answer: Zeroing out negative gradients at ReLU gates during the backward pass

Guided backpropagation zeros gradients that are negative either in the upstream gradient OR in the forward activation, producing sharper, cleaner saliency maps.

Question 110: What is the 'objectness score' in YOLO detections?

The confidence that a bounding box contains any object (Correct answer)
The probability that a detected region belongs to a specific class
The anchor box scale multiplier
The IoU threshold used during NMS

Correct answer: The confidence that a bounding box contains any object

The objectness score estimates the probability that the bounding box actually contains an object (regardless of class).

Question 111: What is the key architectural difference between one-stage and two-stage object detectors?

Two-stage detectors generate region proposals first, then classify them (Correct answer)
Two-stage detectors skip NMS post-processing
One-stage detectors use larger backbones
One-stage detectors only work on small images

Correct answer: Two-stage detectors generate region proposals first, then classify them

Two-stage detectors (e.g., Faster R-CNN) first generate proposals then classify each; one-stage detectors (e.g., YOLO) predict boxes and classes simultaneously.

Question 112: A data scientist wants to understand which specific pixels in an input image are most influential in causing a CNN to make a particular classification decision (e.g., classifying an image as a 'dog'). Which visualization technique would be most appropriate for this purpose?

Activation Maximization.
Saliency Maps (or Gradient-based Attribution). (Correct answer)
t-SNE projection of the final feature vector.
Visualizing the filter weights directly.

Correct answer: Saliency Maps (or Gradient-based Attribution).

Saliency maps are designed to solve this exact problem. They work by computing the gradient of the output class score with respect to the input image pixels. The magnitude of the gradient for each pixel indicates how much a small change in that pixel's intensity would affect the class score, thus highlighting the most influential pixels for that specific classification decision.

Question 113: Which of the following is NOT a common technique used to combat overfitting in a CNN?

Dropout.
Data Augmentation.
Increasing the number of epochs indefinitely. (Correct answer)
L2 Regularization (Weight Decay).

Correct answer: Increasing the number of epochs indefinitely.

Increasing the number of epochs indefinitely is likely to cause overfitting, not prevent it. As a model trains for longer, it has more opportunities to memorize the noise and specific details of the training data. [24] Data augmentation, Dropout, and L2 regularization are all standard and effective techniques used to reduce overfitting and improve a model's generalization capabilities. [12, 20]

Question 114: In a CNN, what does 'stride' control?

The learning rate during backpropagation
The depth of the convolutional filter
The number of channels in the output
How many pixels the filter moves between applications (Correct answer)

Correct answer: How many pixels the filter moves between applications

Stride determines the step size of the filter as it slides across the input, with larger strides producing smaller output feature maps.

Question 115: Which regularization technique randomly zeroes out feature maps during training in CNNs rather than individual neurons?

Weight decay
Spatial Dropout (Correct answer)
Batch Normalization
L2 regularization

Correct answer: Spatial Dropout

Spatial Dropout drops entire feature maps (channels) instead of individual activations, which is more effective for spatially correlated CNN features.

Question 116: Spatial Pyramid Pooling (SPP) enables CNNs to accept inputs of varying sizes primarily by doing what?

Dynamically adjusting kernel sizes in convolutional layers based on input size
Applying pooling at multiple scales and concatenating outputs into a fixed-length vector (Correct answer)
Using dilated convolutions to match arbitrary input resolutions
Resizing all inputs to a fixed dimension before the first convolution

Correct answer: Applying pooling at multiple scales and concatenating outputs into a fixed-length vector

SPP pools the final feature maps at several granularities (e.g., 1×1, 2×2, 4×4) and concatenates the results, always producing a fixed-size representation regardless of input size.

Question 117: Why is it problematic to use the test set to choose the best regularization hyperparameters in a CNN experiment?

It prevents overfitting on training data
It reduces the number of available training samples
It causes information leakage, making the model appear to generalize better than it truly does (Correct answer)
It makes training slower

Correct answer: It causes information leakage, making the model appear to generalize better than it truly does

Using the test set for hyperparameter selection causes data leakage, resulting in over-optimistic performance estimates that do not reflect true generalization.

Question 118: During inference, how does batch normalization behave differently compared to training?

It disables all normalization and passes activations unchanged
It uses population statistics (running mean and variance) instead of batch statistics (Correct answer)
It uses a larger batch to compute more accurate statistics
It applies normalization only to the first and last layers

Correct answer: It uses population statistics (running mean and variance) instead of batch statistics

During inference, batch normalization uses fixed population statistics (running mean and variance accumulated during training) rather than computing batch statistics, ensuring consistent outputs for single samples.

Question 119: In a CNN used for multi-class image classification, which activation function is typically applied at the final output layer?

Softmax (Correct answer)
Sigmoid
Tanh
ReLU

Correct answer: Softmax

Softmax converts raw logits into a probability distribution over all classes, with each output representing the probability that the input belongs to that class.

Question 120: What does the 'momentum' hyperparameter do in SGD with momentum for CNN training?

Scales the learning rate adaptively per parameter
Accumulates a fraction of past gradients to accelerate optimization (Correct answer)
Clips gradients to a maximum norm
Adds L2 regularization to the loss

Correct answer: Accumulates a fraction of past gradients to accelerate optimization

Momentum accumulates an exponentially decaying moving average of past gradients, helping accelerate SGD in relevant directions and dampening oscillations.

Question 121: Which loss function is most appropriate when training a CNN for multi-label classification where each image can belong to multiple classes?

Mean squared error over class logits
Binary cross-entropy with sigmoid applied per class (Correct answer)
Hinge loss with one-vs-all strategy
Categorical cross-entropy with softmax

Correct answer: Binary cross-entropy with sigmoid applied per class

Binary cross-entropy with a sigmoid activation per output neuron treats each class independently as a binary decision, correctly handling multi-label scenarios where multiple classes can be true simultaneously.

Question 122: What role does Batch Normalization play in CNN training?

It randomly drops entire feature maps to improve generalization
It clips gradient values to prevent exploding gradients
It rescales weight matrices to unit norm after each update
It normalizes layer inputs per mini-batch, stabilizing and accelerating training (Correct answer)

Correct answer: It normalizes layer inputs per mini-batch, stabilizing and accelerating training

Batch Normalization normalizes activations to zero mean and unit variance per mini-batch, reducing internal covariate shift and allowing higher learning rates.

Question 123: What is the Pyramid Scene Parsing Network (PSPNet) and what problem does it address?

A network that parses 3D point clouds
A network for parsing text scenes
A segmentation network that uses a Pyramid Pooling Module to capture global and local context at multiple scales (Correct answer)
A recurrent network for sequential scene understanding

Correct answer: A segmentation network that uses a Pyramid Pooling Module to capture global and local context at multiple scales

PSPNet's Pyramid Pooling Module aggregates features at four different scales using adaptive average pooling, enabling the network to incorporate global context which reduces misclassification from local ambiguity.

Question 124: An engineer is choosing a classic CNN architecture for a task on hardware with limited computational resources. They need a model that is deep and accurate but has a significantly lower parameter count compared to VGG-16 or AlexNet. Which architecture would be the most suitable choice?

LeNet-5
GoogLeNet (Inception v1) (Correct answer)
AlexNet
VGG-19

Correct answer: GoogLeNet (Inception v1)

GoogLeNet (Inception v1) was designed for computational efficiency. Through the heavy use of 1x1 convolutions for dimensionality reduction within its Inception modules, it achieved state-of-the-art accuracy with only about 6-7 million parameters, a fraction of AlexNet's (~60 million) and VGGNet's (~138 million) parameters.

Question 125: In a bottleneck residual block (ResNet-50+), what is the purpose of the 1×1 convolutions flanking the 3×3 layer?

To add non-linearity before the skip connection
To reduce and then restore channel depth, lowering computation for the 3×3 step (Correct answer)
To perform batch normalization without an extra layer
To apply spatial attention across the feature map

Correct answer: To reduce and then restore channel depth, lowering computation for the 3×3 step

The flanking 1×1 convolutions compress channels before the expensive 3×3 convolution and expand them back, making the block computationally efficient.

Question 126: What is CutMix augmentation and how does it differ from Mixup?

CutMix applies to text data only
CutMix is the same as Mixup
CutMix pastes a cut-out patch from one image onto another, mixing labels proportionally to the patch area (Correct answer)
CutMix only changes colors

Correct answer: CutMix pastes a cut-out patch from one image onto another, mixing labels proportionally to the patch area

CutMix replaces a rectangular region in one image with the corresponding region from another and mixes labels proportionally to the replaced area, unlike Mixup which blends entire images.

Question 127: What is the approximate top-5 error rate AlexNet achieved on ImageNet ILSVRC 2012?

10.9%
5.0%
25.8%
15.3% (Correct answer)

Correct answer: 15.3%

AlexNet achieved a top-5 error of 15.3%, which was dramatically better than the runner-up's 26.2%, demonstrating deep CNN superiority.

Question 128: ResNet's skip connections primarily help transfer learning by:

Reducing the number of trainable parameters
Skipping layers during inference for speed
Preventing vanishing gradients and enabling very deep pretrained models (Correct answer)
Connecting source and target domain layers

Correct answer: Preventing vanishing gradients and enabling very deep pretrained models

Skip connections allow gradients to flow directly through the network, enabling training of very deep ResNets that provide rich pretrained features.

Question 129: A data scientist is evaluating their object detection model's performance. They are using Intersection over Union (IoU) to determine if a predicted bounding box is a true positive. What does an IoU score of 0.8 signify?

There is a high degree of overlap, with the area of intersection being 80% of the area of the union between the predicted and ground-truth boxes. (Correct answer)
The predicted box and the ground-truth box have no overlap.
The area of the union of the two boxes is 80% of the area of their intersection.
The model is 80% confident that the object class is correct.

Correct answer: There is a high degree of overlap, with the area of intersection being 80% of the area of the union between the predicted and ground-truth boxes.

Intersection over Union (IoU) is a metric used to evaluate the accuracy of a predicted bounding box by measuring how much it overlaps with the ground-truth box. It is calculated as the area of the intersection of the two boxes divided by the area of their union. An IoU score of 0.8 indicates a very good localization, as it means the shared area (intersection) is 80% of the total area covered by both boxes combined (union), signifying a strong overlap.

Question 130: What is the SegNet architecture and what distinguishes its decoder from other segmentation networks?

A network combining segmentation with depth estimation
A segmentation network whose decoder uses pooling indices from the encoder's max-pooling layers to upsample feature maps without learning upsampling weights (Correct answer)
A network using attention for segmentation
A recurrent segmentation network for video

Correct answer: A segmentation network whose decoder uses pooling indices from the encoder's max-pooling layers to upsample feature maps without learning upsampling weights

SegNet's decoder reuses the max-pooling indices (locations of selected values) stored by the encoder to perform non-linear upsampling, reducing memory requirements while preserving spatial structure.

Question 131: In the context of CNNs, what does 'DropBlock' improve upon compared to standard Dropout?

It applies dropout only during backpropagation
It uses a higher drop rate than standard Dropout
It applies dropout only to the final layer
It drops contiguous blocks of spatial units, preventing nearby units from compensating for dropped ones (Correct answer)

Correct answer: It drops contiguous blocks of spatial units, preventing nearby units from compensating for dropped ones

DropBlock removes contiguous regions of feature maps so that semantic information in those regions is completely lost, making regularization more effective for CNNs.

Question 132: Global Average Pooling (GAP) is frequently used at the end of a CNN to replace which traditional layer?

Dropout layer
Batch normalization layer
Fully connected (dense) layer (Correct answer)
Convolutional layer

Correct answer: Fully connected (dense) layer

GAP reduces each feature map to a single value, replacing large fully connected layers and dramatically reducing parameter count.

Question 133: What loss function is most commonly used as the primary training objective in semantic segmentation CNNs?

Triplet Loss
Hinge Loss
Mean Squared Error (MSE)
Pixel-wise Cross-Entropy Loss (Correct answer)

Correct answer: Pixel-wise Cross-Entropy Loss

Pixel-wise cross-entropy treats each pixel as an independent classification problem and sums the cross-entropy loss over all pixels, making it the standard choice for semantic segmentation training.

Question 134: What is the main purpose of using the ReLU (Rectified Linear Unit) activation function in the hidden layers of a CNN?

To significantly increase the number of parameters for better feature extraction.
To convert the output into a probability distribution for classification.
To introduce non-linearity, allowing the network to learn more complex patterns. (Correct answer)
To normalize the output of the layer to have a mean of zero.

Correct answer: To introduce non-linearity, allowing the network to learn more complex patterns.

The ReLU activation function introduces non-linearity into the network. Without a non-linear activation function, a deep CNN would behave like a single, equivalent convolutional layer, limiting its ability to learn complex relationships in the data. ReLU is computationally efficient and helps mitigate the vanishing gradient problem.

Question 135: In the context of pooling, what does 'stride' control?

How much the pooling window moves at each step across the feature map (Correct answer)
The depth of the pooling operation across channels
The activation function applied after pooling
The number of distinct pooling operations applied in parallel

Correct answer: How much the pooling window moves at each step across the feature map

Stride determines the step size of the pooling window as it slides across the feature map, controlling output size and overlap.

Question 136: How does the Feature Pyramid Network (FPN) benefit CNN-based segmentation models?

It applies pyramid-shaped attention masks
It builds a multi-scale feature pyramid by combining high-resolution low-level features with semantically rich high-level features via top-down pathways (Correct answer)
It reduces the number of classes in the output
It reduces training time by skipping intermediate layers

Correct answer: It builds a multi-scale feature pyramid by combining high-resolution low-level features with semantically rich high-level features via top-down pathways

FPN creates a feature pyramid where each level has strong semantics by merging high-level features (downsampled path) with high-resolution features (bottom-up path), enabling detection and segmentation at multiple scales.

Question 137: For more than 30 years, renal replacement therapy (RRT) has been in use. Which of the following is accurate for the approximate mortality rate of the patient with ARF (acute renal failure) despite the time this therapy has been used?

90%
50% (Correct answer)
10%
30%

Correct answer: 50%

Despite significant advancements in renal replacement therapy (RRT) over several decades, the mortality rate for patients experiencing acute renal failure (ARF) remains remarkably high. Clinical studies and data consistently indicate that approximately 50% of patients, particularly those with severe ARF or in critical care settings, do not survive, underscoring the serious nature and challenges in treating this condition.

Question 138: How does the Softmax activation function differ from Sigmoid when used in output layers?

Softmax outputs sum to 1 across all classes, while Sigmoid outputs each class independently (Correct answer)
Softmax introduces sparsity, while Sigmoid ensures dense probability distributions
Softmax clips values above 1, while Sigmoid allows unbounded positive outputs
Softmax is used for binary tasks, while Sigmoid handles multi-class problems

Correct answer: Softmax outputs sum to 1 across all classes, while Sigmoid outputs each class independently

Softmax normalizes outputs so they form a probability distribution summing to 1, making it ideal for mutually exclusive multi-class classification.

Question 139: When applying transfer learning to object detection (e.g., using Faster R-CNN), what part of the network is typically pretrained?

The classification head for bounding boxes
The backbone CNN used as a feature extractor (Correct answer)
The region proposal network (RPN)
The non-maximum suppression module

Correct answer: The backbone CNN used as a feature extractor

The backbone CNN (e.g., ResNet) is pretrained on ImageNet and used to extract feature maps, while detection-specific heads are trained from scratch.

Question 140: Which backbone architecture is commonly used in modern object detectors like Faster R-CNN and RetinaNet?

AlexNet
VGG-7
LeNet-5
ResNet with FPN (Correct answer)

Correct answer: ResNet with FPN

ResNet combined with a Feature Pyramid Network is a widely used backbone for object detection due to its strong multi-scale feature representation.

Question 141: In the Inception module (GoogLeNet), why are convolutions of different kernel sizes applied in parallel?

To reduce the number of training epochs needed
To capture features at multiple scales simultaneously within the same layer (Correct answer)
To eliminate the need for pooling layers
To enforce weight sharing across different resolution branches

Correct answer: To capture features at multiple scales simultaneously within the same layer

Parallel branches with 1×1, 3×3, and 5×5 filters let the Inception module learn features at multiple spatial scales and concatenate them along the depth axis.

Question 142: What distinguishes a depthwise separable convolution from a standard convolution?

It replaces pooling layers entirely
It uses larger kernel sizes to capture more context
It applies one filter per input channel, then combines with 1×1 convolutions (Correct answer)
It operates only on the fully connected layers

Correct answer: It applies one filter per input channel, then combines with 1×1 convolutions

Depthwise separable convolution factorizes a standard convolution into a depthwise step (per-channel) and a pointwise 1×1 step, reducing computation.

Question 143: In the context of CNNs, what is the main advantage of using batch normalization over dropout for regularization in convolutional layers?

Batch normalization can be applied to the input images, while dropout cannot
Batch normalization eliminates the need for weight initialization strategies
Batch normalization provides regularization while also stabilizing and accelerating training, without discarding information (Correct answer)
Batch normalization uses fewer memory resources than dropout during training

Correct answer: Batch normalization provides regularization while also stabilizing and accelerating training, without discarding information

Batch normalization regularizes convolutional layers by reducing internal covariate shift while simultaneously speeding convergence, whereas dropout discards activation values and can harm feature learning in convolutional layers.

Question 144: Why does soft-NMS outperform standard NMS in crowded scene detection?

It removes all overlapping boxes regardless of score
It uses a learned threshold instead of a fixed IoU cutoff
It decays scores of overlapping boxes rather than eliminating them (Correct answer)
It applies NMS only to the highest confidence class

Correct answer: It decays scores of overlapping boxes rather than eliminating them

Soft-NMS reduces suppressed boxes' scores by a continuous function of overlap instead of hard removal, preserving detections of nearby objects.

Question 145: What does 'dilation' (atrous convolution) add to a standard convolution?

Spaces between filter elements, enlarging the receptive field without increasing parameters (Correct answer)
A skip connection that bypasses the current layer
Additional convolutional channels at each layer
Extra zero-padding around the input border

Correct answer: Spaces between filter elements, enlarging the receptive field without increasing parameters

Dilated convolution inserts gaps (holes) between filter weights, exponentially growing the receptive field while keeping the parameter count and resolution constant.

Question 146: Which classic CNN architecture introduced the concept of training two parallel GPU pipelines (splitting feature maps across GPUs)?

VGGNet
AlexNet (Correct answer)
LeNet-5
GoogLeNet

Correct answer: AlexNet

AlexNet split its feature maps across two GPUs to fit within the 3GB VRAM of 2012 GTX 580 GPUs, an engineering constraint that shaped its architecture.

Question 147: What is the purpose of the squeeze-and-excitation (SE) block in SENet?

To merge feature maps from different layers using element-wise addition
To reduce spatial resolution by 50% between convolutional blocks
To recalibrate channel-wise feature responses by modeling inter-channel dependencies (Correct answer)
To apply group normalization across subsets of channels

Correct answer: To recalibrate channel-wise feature responses by modeling inter-channel dependencies

The SE block globally pools feature maps, learns channel importance weights via two FC layers, and scales channels accordingly — acting as channel-wise attention.

Question 148: How many parameters does GoogLeNet have compared to AlexNet?

GoogLeNet has roughly 12x more parameters
GoogLeNet has roughly 4x fewer parameters
GoogLeNet has roughly the same number of parameters
GoogLeNet has roughly 12x fewer parameters (Correct answer)

Correct answer: GoogLeNet has roughly 12x fewer parameters

AlexNet has ~60M parameters while GoogLeNet has only ~5M, making GoogLeNet roughly 12x more parameter-efficient due to its Inception modules.

Question 149: A CNN feature map of size 8×8 is processed by 2×2 max pooling with stride 2 applied twice sequentially. What is the final output size?

4×4
6×6
2×2 (Correct answer)
1×1

Correct answer: 2×2

After the first pooling: 8→4; after the second pooling: 4→2, so the final feature map is 2×2.

Question 150: Which architecture introduced the concept of 'network-in-network' using micro neural networks at each convolutional step?

SqueezeNet
AlexNet
Network-in-Network (NIN) (Correct answer)
VGGNet

Correct answer: Network-in-Network (NIN)

NIN replaced standard convolutional filters with small multi-layer perceptrons (mlpconv layers) at each location, enabling more complex feature extraction per patch.

Certified Nephrology Nurse Exam

The CNN exam, administered by the Nephrology Nursing Certification Commission (NNCC), is a computer-based test consisting of 150 questions (130 scored, 20 unscored pretest) with a 3-hour time limit. A passing standard score of 95 is required, equivalent to answering approximately 70% of scored questions correctly. Content spans five clinical areas: concepts of kidney disease, hemodialysis, peritoneal dialysis, transplant, and acute therapies, along with professional practice standards.

150

Questions

180 min

Time Limit

70%

To Pass

Checking for saved progress...

Exam Rules

You can skip questions and return to them later
Flag questions for review before submitting
No feedback shown until you submit the entire exam
Unanswered questions count as wrong — answer everything
10 pretest questions are mixed in and don't affect your score
Timer auto-submits when time runs out
Your progress is auto-saved every 30 seconds