Certified Nephrology Nurse Exam — Questions and Answers

Question 1: What is the main purpose of using the ReLU (Rectified Linear Unit) activation function in the hidden layers of a CNN?

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

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

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
To create an ensemble of predictions at inference

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

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

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 4: 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?

A fully connected (dense) layer (Correct answer)
A max-pooling layer
Another convolutional 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 5: What is 'cyclical learning rates' (CLR) in CNN optimization?

Alternating between SGD and Adam optimizers
Oscillating the learning rate between a minimum and maximum bound in cycles (Correct answer)
Using different learning rates for each layer
Decreasing the learning rate exponentially across training

Correct answer: Oscillating the learning rate between a minimum and maximum bound in cycles

CLR varies the learning rate cyclically between a lower and upper bound, allowing the optimizer to periodically escape saddle points and explore the loss surface more effectively.

Question 6: In AlexNet, what was the purpose of Local Response Normalization (LRN)?

To replace dropout during training
To normalize activations across neighboring feature maps for lateral inhibition (Correct answer)
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 7: The gradient of a substance's concentration, such as urea, during peritoneal dialysis

Increases
Decreases (Correct answer)
Remains consistent
Varies widely

Correct answer: Decreases

During peritoneal dialysis, the dialysate solution is introduced into the peritoneal cavity, creating a concentration gradient for waste products like urea. Initially, urea concentration is high in the blood and low in the dialysate. As dialysis progresses, urea moves from the blood into the dialysate, causing its concentration in the blood to decrease and its concentration in the dialysate to increase, thus reducing the overall concentration gradient over time until equilibrium is approached.

Question 8: A data scientist is designing a CNN to classify high-resolution medical images where preserving spatial information in the early layers is critical. Which hyperparameter setting for the initial convolutional layers would be most appropriate?

Large filter size (e.g., 11x11) and a large stride (e.g., 4).
A 1x1 filter with a stride of 1.
No padding ('valid') and a stride of 2.
Small filter size (e.g., 3x3), stride of 1, and 'same' padding. (Correct answer)

Correct answer: Small filter size (e.g., 3x3), stride of 1, and 'same' padding.

To preserve spatial information, it is best to avoid aggressive down-sampling in the early layers. A small filter size (3x3) captures local features, a stride of 1 ensures the filter moves pixel-by-pixel without skipping information, and 'same' padding ensures that the output feature map has the same spatial dimensions as the input, preventing information loss at the borders.

Question 9: In a depthwise separable convolution, what does the depthwise step do?

Applies one filter across all channels
Combines all channels into one output
Performs 1×1 convolutions only
Applies a separate filter to each input channel independently (Correct answer)

Correct answer: Applies a separate filter to each input channel independently

The depthwise step convolves each input channel with its own dedicated spatial filter, keeping channels separate.

Question 10: A CNN's training curve shows decreasing training loss but a validation loss that oscillates without clear improvement. What does this indicate?

The model is overfitting and the validation set may have high variance (Correct answer)
The learning rate is too low
Gradient vanishing is present
The model is underfitting and needs more capacity

Correct answer: The model is overfitting and the validation set may have high variance

Decreasing training loss alongside flat or oscillating validation loss is a sign of overfitting, possibly compounded by a small or noisy validation set.

Question 11: What is the 'dead ReLU' problem during CNN training?

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

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 12: In PyTorch, how do you freeze all parameters of a pretrained model before adding a new classifier?

model.eval()
model.freeze()
torch.no_grad()
for param in model.parameters(): param.requires_grad = False (Correct answer)

Correct answer: for param in model.parameters(): param.requires_grad = False

Setting requires_grad = False on all parameters prevents gradient computation and weight updates for the pretrained backbone.

Question 13: The LeNet-5 architecture, a pioneering CNN, was primarily designed for a specific task and had characteristic features for its time. Which of the following best describes LeNet-5?

A shallow network using average pooling and tanh/sigmoid activations, designed for handwritten digit recognition. (Correct answer)
A model that introduced residual connections to train networks over 100 layers deep.
An architecture that used parallel multi-scale convolutions and global average pooling.
A deep 19-layer network using only 3x3 filters for large-scale color image classification.

Correct answer: A shallow network using average pooling and tanh/sigmoid activations, designed for handwritten digit recognition.

LeNet-5, developed by Yann LeCun, was a foundational 7-layer CNN designed for handwritten digit recognition on the MNIST dataset. Its architecture was characterized by its relatively shallow depth, alternating convolutional and average pooling layers, and the use of sigmoid or tanh activation functions.

Question 14: What are the two learnable parameters introduced by batch normalization?

Weight decay (lambda) and dropout rate (p)
Scale (gamma) and shift (beta) (Correct answer)
Learning rate (alpha) and momentum (beta)
Mean (mu) and variance (sigma)

Correct answer: Scale (gamma) and shift (beta)

Batch normalization introduces gamma (scale) and beta (shift) parameters that allow the network to undo the normalization if needed, learned during training.

Question 15: In CNN training, what does the term 'augmentation policy' refer to?

A defined set of transformation operations along with their probabilities and magnitudes (Correct answer)
The model regularization method
The GPU policy for running training
The learning rate schedule

Correct answer: A defined set of transformation operations along with their probabilities and magnitudes

An augmentation policy specifies which transformations to apply, with what probability each is chosen, and at what intensity level, forming a complete augmentation strategy.

Question 16: 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?

Mixup
Mosaic
RandomCrop
CutMix (Correct answer)

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

Sparse connectivity
Weight sharing (parameter sharing) (Correct answer)
Translation invariance
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 18: 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?

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

Correct answer: Catastrophic forgetting of pretrained features

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

Question 19: During CNN training, which data augmentation technique generates new training samples by combining two images and their labels linearly?

Random Erasing
Mixup (Correct answer)
CutMix
Cutout

Correct answer: Mixup

Mixup creates training samples as convex combinations of pairs of training examples and their labels.

Question 20: Which Python library function is commonly used to extract intermediate layer outputs for feature map visualization in Keras?

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

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

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

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

To apply group normalization across subsets of channels
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)

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 23: What is the key computational advantage of max pooling over learned downsampling (e.g., strided convolutions)?

Max pooling has no trainable parameters, reducing model complexity and memory (Correct answer)
Max pooling introduces additional trainable weights that improve accuracy
Max pooling applies batch normalization implicitly during downsampling
Max pooling increases feature map depth, capturing more information

Correct answer: Max pooling has no trainable parameters, reducing model complexity and memory

Max pooling is a fixed operation with no learnable parameters, making it cheaper in memory and compute compared to strided convolutions that add weights.

Question 24: What is the approximate top-5 error rate AlexNet achieved on ImageNet ILSVRC 2012?

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

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 25: How does RetinaNet address the foreground-background class imbalance problem?

By using hard negative mining
By training with balanced mini-batches
By applying Focal Loss (Correct answer)
By adding a separate background classifier

Correct answer: By applying Focal Loss

RetinaNet uses Focal Loss, which down-weights the contribution of easy negatives so training focuses on hard examples.

Question 26: What is the purpose of transposed convolutions (deconvolutions) in segmentation networks?

To reduce the number of feature map channels
To apply convolutions in the reverse order
To learn to upsample feature maps back to a higher spatial resolution (Correct answer)
To transpose the weight matrix of convolution layers

Correct answer: To learn to upsample feature maps back to a higher spatial resolution

Transposed convolutions perform a learnable upsampling by inserting zeros between input values and then applying a convolution, allowing the network to recover spatial resolution lost during downsampling.

Question 27: How does increasing filter size (e.g., from 3×3 to 7×7) affect the receptive field of a single convolutional layer?

Larger receptive field per layer (Correct answer)
Smaller receptive field per layer
Receptive field becomes equal to input size
No change to receptive field

Correct answer: Larger receptive field per layer

A larger filter directly covers more spatial area of the input, increasing the receptive field of each output neuron in that layer.

Question 28: What effect does L1 regularization have on CNN weight vectors compared to L2?

It doubles the gradient magnitude
It produces uniformly small weights
It increases the learning rate
It produces sparse weights with many zeros (Correct answer)

Correct answer: It produces sparse weights with many zeros

L1 regularization adds an absolute-value penalty that drives many weights exactly to zero, producing sparse weight vectors.

Question 29: Which backbone architecture is commonly used in modern object detectors like Faster R-CNN and RetinaNet?

LeNet-5
VGG-7
AlexNet
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 30: Which of the following best describes the role of the convolutional layer in a CNN?

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

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 31: What problem does gradient checkpointing solve in CNN training?

Slow convergence due to poor initialization
Vanishing gradients in deep networks
Exploding gradients during training
High memory usage during backpropagation by recomputing activations (Correct answer)

Correct answer: High memory usage during backpropagation by recomputing activations

Gradient checkpointing reduces memory by not storing all intermediate activations; instead, it recomputes them during the backward pass, trading compute time for memory savings.

Question 32: What is the receptive field of a neuron in a CNN?

The region of the input image that influences that neuron's activation (Correct answer)
The number of channels it processes
The output size of the layer it belongs to
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 33: 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 acts as a pooling operation
It changes the number of channels without altering spatial dimensions (Correct answer)

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 34: What is 'task-specific fine-tuning' in contrast to 'feature extraction' transfer learning?

Using task-specific data only for augmentation
Updating pretrained weights (some or all) for the new task (Correct answer)
Training only the softmax layer on the new task
Extracting features specific to the source task

Correct answer: Updating pretrained weights (some or all) for the new task

Task-specific fine-tuning allows pretrained weights to be updated via backpropagation on the new task, unlike feature extraction which keeps them frozen.

Question 35: Which loss function is most appropriate when training a CNN for multi-label classification where each image can belong to multiple classes?

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

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

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

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

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

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

The number of filters applied to the input volume.
The amount of zero-padding added to the borders of the input.
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.

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 39: In YOLO, what does each grid cell predict?

Bounding boxes, objectness scores, and class probabilities (Correct answer)
Region proposals for the RPN
Only the bounding box coordinates
Only the class probabilities

Correct answer: Bounding boxes, objectness scores, and class probabilities

Each YOLO grid cell predicts B bounding boxes with confidence scores and C class probabilities for objects whose center falls in that cell.

Question 40: A researcher is working with a standard ResNet-50 model and wants to generate a class-specific heatmap showing important image regions for a prediction. Since this architecture doesn't end with a Global Average Pooling (GAP) layer suitable for classic CAM, which popular technique can they use without retraining or modifying the model?

Principal Component Analysis (PCA) on feature maps.
Gradient-weighted Class Activation Mapping (Grad-CAM). (Correct answer)
Deconvolution (Transposed Convolution).
Layer-wise Relevance Propagation (LRP).

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

Grad-CAM is a generalization of CAM designed specifically for this scenario. It overcomes the architectural limitation of requiring a GAP layer by using the gradient information flowing into the final convolutional layer. These gradients are used to weight the feature maps, indicating their importance for a specific class prediction, making it applicable to a wide range of CNN architectures like ResNet without modification.

Question 41: What is the 'dying ReLU' problem?

Neurons with ReLU permanently output zero because their weights push inputs negative (Correct answer)
ReLU outputs explode to infinity during training
ReLU saturates at large positive values, halting learning
ReLU causes gradients to grow exponentially

Correct answer: Neurons with ReLU permanently output zero because their weights push inputs negative

When neurons receive consistently negative inputs, ReLU outputs zero with zero gradient, so weights never update—the neuron effectively 'dies.'

Question 42: What is fractional max pooling?

Pooling that averages only the top fraction of activations in a window
A technique combining max and average pooling with a learned weight
Max pooling with a non-integer kernel size or stride, producing irregular output dimensions (Correct answer)
A pooling method that applies different pool sizes per channel

Correct answer: Max pooling with a non-integer kernel size or stride, producing irregular output dimensions

Fractional max pooling uses random or pseudo-random pooling regions with non-integer reduction ratios, acting as a form of data augmentation.

Question 43: Which of the following best describes the primary purpose of a 1x1 convolution operation in a CNN architecture?

To perform spatial down-sampling similar to a pooling layer.
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.

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 44: What happens to dropout during model inference/testing?

A new random mask is sampled and applied once at inference time
Dropout continues but only in the final fully connected layer
All neurons are active and outputs are scaled by the keep probability (Correct answer)
Dropout is applied with half the training dropout rate

Correct answer: All neurons are active and outputs are scaled by the keep probability

At inference time, dropout is disabled and all neurons are active; outputs are typically scaled by the keep probability (1 - dropout rate) to match expected values during training.

Question 45: Dialysate glucose must be heat sterilized at a low pH to

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

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 46: 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 apply spatial attention across the feature map
To reduce and then restore channel depth, lowering computation for the 3×3 step (Correct answer)
To perform batch normalization without an extra layer

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 47: Which optimizer introduces a per-parameter adaptive learning rate based on the second moment of gradients?

SGD with momentum
Adam (Correct answer)
RMSprop
Adagrad

Correct answer: Adam

Adam maintains both first moment (mean) and second moment (uncentered variance) estimates of gradients, providing adaptive per-parameter learning rates with bias correction.

Question 48: What is the 'class imbalance problem' that one-stage detectors face during training?

The model learns faster for large objects than small ones
Some object classes are harder to classify than others
Different feature map scales have unequal numbers of anchor boxes
There are vastly more background anchor boxes than foreground object anchors (Correct answer)

Correct answer: There are vastly more background anchor boxes than foreground object anchors

In one-stage detectors, the majority of anchor boxes cover background, creating a massive imbalance that can dominate the loss and degrade training.

Question 49: What information does a 'mean activation map' (averaging all channels of a feature tensor) convey?

A spatial summary of overall network attention — where the layer as a whole responds strongly (Correct answer)
The class probability at each spatial location
The gradient magnitude at each position
Which filters have the largest weights

Correct answer: A spatial summary of overall network attention — where the layer as a whole responds strongly

Averaging across all channels of a feature tensor collapses filter-specific detail into a single spatial map showing the regions that collectively activate most across all detectors.

Question 50: What does batch normalization normalize during training?

The weights of each convolutional filter
The learning rate at each iteration
The gradient magnitudes before backpropagation
The activations of each layer across a mini-batch (Correct answer)

Correct answer: The activations of each layer across a mini-batch

Batch normalization normalizes the activations (outputs) of a layer across the current mini-batch to have zero mean and unit variance.

Question 51: What is multi-task learning in the context of CNN transfer learning?

Training a single CNN on multiple related tasks simultaneously to share representations (Correct answer)
Using multiple optimizers for different layers
Sequentially fine-tuning on different datasets
Running multiple pretrained models in parallel

Correct answer: Training a single CNN on multiple related tasks simultaneously to share representations

Multi-task learning shares a CNN backbone across multiple tasks, allowing shared representations to improve generalization on each task.

Question 52: A developer is building a flower species classifier using a small, custom dataset of approximately 900 images. They decide to use a ResNet50 model pre-trained on ImageNet. To leverage the pre-trained features effectively while minimizing the risk of overfitting, which transfer learning strategy should they implement first?

Unfreeze all layers and fine-tune the entire network using a high learning rate.
Remove the first few convolutional layers and retrain the rest of the network on the new dataset.
Freeze all convolutional layers and train only a new, randomly initialized classifier head. (Correct answer)
Train the entire network from scratch with randomly initialized weights.

Correct answer: Freeze all convolutional layers and train only a new, randomly initialized classifier head.

With a small dataset, there is a high risk of overfitting if the entire network is trained. The best initial strategy is feature extraction, which involves freezing the pre-trained convolutional layers to use their powerful, generic feature detection capabilities and only training the new classifier head on the small dataset.

Question 53: When stride > 1 is used in a convolutional layer, what is the primary effect?

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

Correct answer: Spatial downsampling of the feature map

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

Question 54: What is 'negative transfer' in the context of transfer learning with CNNs?

Transferring weights with negative values
When pretraining hurts performance on the target task (Correct answer)
Transfer learning that reduces model size
Training on negative examples only

Correct answer: When pretraining hurts performance on the target task

Negative transfer occurs when features from the source domain are incompatible with the target domain, degrading target task performance.

Question 55: Which of the following is a defining characteristic of one-stage object detectors like SSD (Single Shot MultiBox Detector)?

They first generate region proposals and then pass each proposal to a separate classifier.
They require each image to be processed multiple times to find all objects.
They use a computationally expensive selective search to find interesting regions first.
They make predictions on a dense grid of locations across feature maps of multiple scales. (Correct answer)

Correct answer: They make predictions on a dense grid of locations across feature maps of multiple scales.

A key feature of one-stage detectors like SSD is that they bypass the explicit region proposal step found in two-stage detectors. Instead, they apply a set of convolutional filters to feature maps at different scales in the network. At each location on these feature maps (a dense grid), the model predicts bounding box offsets and class probabilities simultaneously. Using multiple feature maps allows the model to detect objects of various sizes in a single forward pass.

Question 56: In a CNN, what does 'stride' control?

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

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 57: What architecture is U-Net primarily designed for, and what is its signature structural feature?

Video segmentation with temporal convolutions
Object detection with anchor boxes
Image classification with residual blocks
Biomedical image segmentation with a symmetric encoder-decoder and skip connections (Correct answer)

Correct answer: Biomedical image segmentation with a symmetric encoder-decoder and skip connections

U-Net is a symmetric encoder-decoder network with skip connections that directly concatenate feature maps from the encoder to the corresponding decoder level, preserving fine-grained spatial detail.

Question 58: A CNN is being designed to classify images where the most prominent feature (e.g., a bright edge) in a local region is the most important for classification. Which pooling strategy would be most appropriate?

Average Pooling
Global Pooling
Max Pooling (Correct answer)
Min Pooling

Correct answer: Max Pooling

Max pooling selects the maximum value from each patch of the feature map. This is particularly effective at capturing the most prominent or intense features, such as bright edges or corners, while discarding less relevant information.

Question 59: When visualizing early-layer feature maps, what kind of patterns are typically detected?

Class probability distributions
Abstract semantic concepts
Complex object parts like eyes or wheels
Low-level patterns like edges, colors, and textures (Correct answer)

Correct answer: Low-level patterns like edges, colors, and textures

Early CNN layers respond to simple, low-level visual features such as oriented edges, color gradients, and basic textures.

Question 60: 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 uses fewer memory resources than dropout during training
Batch normalization provides regularization while also stabilizing and accelerating training, without discarding information (Correct answer)
Batch normalization eliminates the need for weight initialization strategies

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

AlexNet (Correct answer)
GoogLeNet
LeNet-5
VGGNet

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

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

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

The region proposal network (RPN)
The classification head for bounding boxes
The backbone CNN used as a feature extractor (Correct answer)
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 64: Why might feature maps of a batch-normalized CNN layer look different when the model is in training mode vs. inference mode?

The learning rate affects activation values directly
Batch norm uses mini-batch statistics during training but running mean/variance during inference, changing activation distributions (Correct answer)
Dropout is applied to feature maps only during training
The filter weights are updated only during training mode

Correct answer: Batch norm uses mini-batch statistics during training but running mean/variance during inference, changing activation distributions

During training, batch normalization normalizes using the current mini-batch's mean and variance; during inference it uses accumulated running statistics, which can shift activation values.

Question 65: 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
LeNet-5
AlexNet
ResNet (Correct answer)

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

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

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 67: A CNN model's training loss is decreasing very slowly, and the model seems to be stuck. To accelerate convergence, the data scientist decides to switch from standard SGD to an adaptive learning rate optimizer. Which optimizer combines the benefits of both momentum and adaptive learning rates by keeping track of exponentially decaying averages of past gradients and past squared gradients?

Adagrad
Nesterov Accelerated Gradient
Adam (Correct answer)
RMSprop

Correct answer: Adam

The Adam optimizer, which stands for Adaptive Moment Estimation, computes adaptive learning rates for each parameter. It stores an exponentially decaying average of past squared gradients like RMSprop, and it also keeps an exponentially decaying average of past gradients, which is similar to momentum. [32, 35] This combination often leads to faster convergence than other optimizers like standard SGD or even RMSprop alone. [34]

Question 68: Why might a practitioner choose strided convolutions over max pooling for downsampling in a modern CNN?

Strided convolutions allow the network to learn the optimal downsampling function rather than using a fixed rule (Correct answer)
Strided convolutions are parameter-free, reducing model size compared to pooling
Strided convolutions always outperform max pooling on benchmark tasks regardless of architecture
Strided convolutions prevent the vanishing gradient problem unlike max pooling

Correct answer: Strided convolutions allow the network to learn the optimal downsampling function rather than using a fixed rule

Strided convolutions have learnable weights, so the network can optimize how it reduces spatial resolution, whereas max pooling applies a fixed non-learned operation.

Question 69: In CNN training, what is 'batch size' effect on generalization — specifically why do smaller batches often generalize better?

Smaller batches converge faster to the global minimum
Smaller batches compute more accurate gradients
Smaller batches introduce more gradient noise acting as implicit regularization (Correct answer)
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 70: What is the primary advantage of using depthwise convolutions over standard convolutions in terms of computational cost?

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)
They require more memory but execute faster on GPU

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

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

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 72: What is the effect of using a very high dropout rate (e.g., 0.9) on a CNN?

The network trains faster because fewer parameters need to be updated
The network becomes more accurate because more neurons are regularized
The batch normalization layers compensate automatically for the high dropout
The network may underfit as too much information is lost during each forward pass (Correct answer)

Correct answer: The network may underfit as too much information is lost during each forward pass

A very high dropout rate removes too many neurons per forward pass, causing the network to lose too much information and potentially underfit — regularization becomes so aggressive it prevents learning.

Question 73: In COCO evaluation, what does AP@[0.5:0.95] represent?

mAP computed only on objects with size between 0.5 and 0.95 of image area
Average Precision averaged over IoU thresholds from 0.5 to 0.95 in 0.05 steps (Correct answer)
The precision at 95% recall
Accuracy at a fixed IoU threshold of 0.5

Correct answer: Average Precision averaged over IoU thresholds from 0.5 to 0.95 in 0.05 steps

COCO's primary metric averages AP across ten IoU thresholds (0.50, 0.55, …, 0.95), rewarding more precise localization than the VOC AP@0.5 metric.

Question 74: Which metric best indicates whether a source domain is suitable for transfer to a target domain?

Batch size used during source training
Domain similarity or distribution overlap (Correct answer)
Depth of the source CNN
Number of parameters in the source model

Correct answer: Domain similarity or distribution overlap

Domain similarity determines how transferable the learned features are; high similarity means better transfer.

Question 75: Which architecture introduced the concept of 'network-in-network' using micro neural networks at each convolutional step?

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

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.

Question 76: What is 'Nesterov Accelerated Gradient' (NAG) and how does it differ from standard momentum SGD?

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

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 77: In progressive fine-tuning, layers are unfrozen from which direction?

From the last (output) layer backward (Correct answer)
Only middle layers first
All layers simultaneously
From the first (input) layer forward

Correct answer: From the last (output) layer backward

Progressive (gradual) fine-tuning unfreezes layers starting from the top (task-specific) toward the bottom (general) to preserve low-level features.

Question 78: 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?

t-SNE projection of the final feature vector.
Saliency Maps (or Gradient-based Attribution). (Correct answer)
Activation Maximization.
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 79: What is 'knowledge distillation' in the context of training CNNs?

Compressing model weights using quantization
Extracting feature representations for unsupervised learning
Training a smaller student model to mimic the outputs of a larger teacher model (Correct answer)
Pruning unimportant neurons from a trained network

Correct answer: Training a smaller student model to mimic the outputs of a larger teacher model

Knowledge distillation trains a compact student CNN to reproduce the soft probability outputs of a larger pretrained teacher, transferring generalization capability to the smaller model.

Question 80: What happens to spatial dimensions when a 5×5 filter is applied with stride=2 and no padding to a 32×32 input?

16×16
28×28
14×14 (Correct answer)
15×15

Correct answer: 14×14

Output size = floor((32 - 5) / 2) + 1 = floor(13.5) + 1 = 14, giving 14×14.

Question 81: What does 'anchor-free' mean in the context of object detectors like FCOS?

Bounding boxes are predicted without predefined anchor boxes (Correct answer)
No backbone network is used
No NMS post-processing is applied
No classification head is used

Correct answer: Bounding boxes are predicted without predefined anchor boxes

Anchor-free detectors like FCOS predict object locations directly from feature map points without relying on predefined anchor boxes.

Question 82: The use of an amino-based dialysate solution is primarily for

Increased ultrafiltration
Nutritional supplementation (Correct answer)
Decreased ultrafiltration
Electrolyte imbalance

Correct answer: Nutritional supplementation

Amino acid-based dialysate solutions are primarily used in peritoneal dialysis for nutritional supplementation. Patients undergoing peritoneal dialysis can experience significant protein loss into the dialysate, leading to malnutrition. By including amino acids in the dialysate, these essential building blocks can be absorbed by the patient, helping to maintain or improve their nutritional status.

Question 83: When applying transfer learning to a new image classification task with a very small and highly similar dataset compared to the original pre-trained model's dataset (e.g., ImageNet), what is the most appropriate strategy?

Use only the first few convolutional layers and build a new, deeper classifier on top.
Re-train the entire network from scratch with a very high learning rate.
Unfreeze all layers and fine-tune the entire network with a moderately high learning rate.
Freeze all convolutional layers and only train the final, newly added classification layers. (Correct answer)

Correct answer: Freeze all convolutional layers and only train the final, newly added classification layers.

When the new dataset is small and similar to the original, the pre-trained convolutional layers already act as excellent generic feature extractors. [8] The best strategy is feature extraction: freeze the base model's layers to prevent overfitting on the small dataset and only train the new classifier head to adapt it to the new set of classes. [1, 23, 28] Re-training all layers (fine-tuning) would risk overfitting, while training from scratch would be ineffective with a small dataset. [23]

Question 84: Which would be the best test to establish whether a person's kidneys operate within normal ranges?

Serum nitrates
Glomerular filtration rate (Correct answer)
Serum glucose
Serum creatinine

Correct answer: Glomerular filtration rate

The Glomerular Filtration Rate (GFR) is widely considered the most accurate and comprehensive measure of overall kidney function. It estimates the volume of blood filtered by the glomeruli per unit of time, directly reflecting the kidneys' ability to remove waste products. While serum creatinine is a component used in GFR calculations, the GFR itself provides a more precise assessment of kidney health than isolated blood markers.

Question 85: Which of the following best describes the 'neck' component in modern object detection architectures like YOLOv4?

The loss computation layer
The final classification head
Feature aggregation module between backbone and detection head (Correct answer)
The input preprocessing pipeline

Correct answer: Feature aggregation module between backbone and detection head

The neck (e.g., PANet, FPN) aggregates and mixes features from different backbone stages before passing them to the detection head.

Question 86: Which transfer learning strategy is best when the target dataset is small and similar to the source dataset?

Use the pretrained model without any modification
Fine-tune all layers with high learning rate
Train from scratch on target data only
Freeze the convolutional base, train only the classifier (Correct answer)

Correct answer: Freeze the convolutional base, train only the classifier

When datasets are similar and target data is limited, freezing the base preserves useful features while avoiding overfitting in the new classifier.

Question 87: What is the effect of applying batch normalization after a convolutional layer before the activation function?

Removes the need for filters entirely
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)

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 88: How does layer normalization differ from batch normalization?

Layer normalization uses a larger batch size to compute more stable statistics
Layer normalization normalizes across the feature/channel dimension for each sample independently, not across the batch (Correct answer)
Layer normalization only normalizes the weight matrices, not the activations
Layer normalization applies normalization only to the final classification layer

Correct answer: Layer normalization normalizes across the feature/channel dimension for each sample independently, not across the batch

Layer normalization normalizes across all features for a single training example rather than across examples in a batch, making it batch-size independent and effective for recurrent networks and transformers.

Question 89: 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
Dropout layers applied at every spatial location
Transposed convolutions (or bilinear upsampling) to recover spatial resolution (Correct answer)

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

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

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

Knowledge distillation
Label smoothing (Correct answer)
Temperature scaling
Mixup

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 92: In a CNN with 64 filters in a convolutional layer, how many feature maps does that layer produce per input image?

1
Depends on input channels
128
64 (Correct answer)

Correct answer: 64

Each filter produces exactly one feature map, so 64 filters yield 64 feature maps per input sample.

Question 93: What is the purpose of 'maximally activating patches' in CNN visualization?

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

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 94: Why are deeper convolutional layers in a CNN said to have larger 'receptive fields'?

Each neuron integrates information from an increasingly larger region of the original input (Correct answer)
They have more channels
They use larger filters than early layers
Their filters are applied with larger strides

Correct answer: Each neuron integrates information from an increasingly larger region of the original input

As feature maps are stacked, each subsequent neuron indirectly covers a larger spatial region of the original input, giving it a larger effective receptive field.

Question 95: What is weight decay in the context of CNN optimization, and how does it relate to L2 regularization?

It is always identical to L1 regularization
It only applies to bias terms
It increases weight magnitudes over time
It is equivalent to L2 regularization for SGD but differs for Adam (Correct answer)

Correct answer: It is equivalent to L2 regularization for SGD but differs for Adam

Weight decay and L2 regularization are equivalent under SGD but diverge under adaptive optimizers like Adam due to interaction with moment estimates.

Question 96: In the Inception module (GoogLeNet), why are convolutions of different kernel sizes applied in parallel?

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

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

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 98: In a convolutional layer, what does 'same' padding ensure?

The filter size matches the input size
Stride is always 1
Output spatial size equals input spatial size (Correct answer)
All padding values are the same number

Correct answer: Output spatial size equals input spatial size

'Same' padding adds zeros around the input so the output height and width match the input when stride=1.

Question 99: Which activation function outputs values strictly between 0 and 1 and is commonly used for binary classification output layers?

Sigmoid (Correct answer)
ReLU
SELU
Tanh

Correct answer: Sigmoid

Sigmoid maps any real-valued input to the (0, 1) range, making it suitable for binary classification where outputs represent probabilities.

Question 100: Feature maps from different channels in the same convolutional layer represent what?

Different learned detectors applied to the same spatial locations (Correct answer)
Different color channels of the output
Separate images in the batch
The same filter applied at different strides

Correct answer: Different learned detectors applied to the same spatial locations

Each channel in a convolutional layer's output corresponds to a different learned filter, so channels capture different types of patterns at the same spatial resolution.

Question 101: Why is local connectivity in convolutional layers advantageous over full connectivity for image data?

Nearby pixels are more correlated, so local filters capture meaningful spatial patterns with fewer parameters (Correct answer)
Full connectivity is computationally cheaper
It removes the need for activation functions
Local connectivity allows larger batch sizes

Correct answer: Nearby pixels are more correlated, so local filters capture meaningful spatial patterns with fewer parameters

Images have strong local spatial correlations, and local filters exploit this structure efficiently while greatly reducing the parameter count.

Question 102: Which pooling operation is most commonly used in CNNs and what does it compute?

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

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 103: EfficientNet scales CNN depth, width, and resolution using a principle called:

Compound scaling (Correct answer)
Residual scaling
Progressive resizing
Layer-wise adaptive rate scaling

Correct answer: Compound scaling

EfficientNet uses compound scaling to uniformly scale all three dimensions (depth, width, resolution) for balanced performance.

Question 104: What is the key innovation in MobileNetV2's inverted residual block compared to standard residual blocks?

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

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

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

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 106: LeNet-5 uses which type of pooling in its subsampling layers?

Fractional pooling
Average pooling (Correct answer)
Max pooling
Global average pooling

Correct answer: Average pooling

LeNet-5 uses average pooling (called 'subsampling') in its pooling layers, which was the standard before max pooling became dominant.

Question 107: A key innovation of the GoogLeNet (Inception v1) architecture was the 'Inception module'. What is the primary purpose of this module?

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

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 108: During the forward pass, what is the bias term in a convolutional layer added to?

The loss function directly
Each input pixel
The filter weights before convolution
Each output feature map value (one bias per filter) (Correct answer)

Correct answer: Each output feature map value (one bias per filter)

One scalar bias per filter is added to every spatial position in that filter's output feature map.

Question 109: What is the DeepLab v3+ architecture's improvement over earlier DeepLab versions?

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

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 110: What distinguishes the ELU (Exponential Linear Unit) activation from ReLU?

ELU outputs negative values for negative inputs using an exponential, giving non-zero mean activations (Correct answer)
ELU applies a sigmoid curve across all input values
ELU clips activations at a fixed maximum value to prevent explosion
ELU uses a linear function for all positive inputs and zero for negatives

Correct answer: ELU outputs negative values for negative inputs using an exponential, giving non-zero mean activations

ELU uses an exponential decay for negative inputs, pushing mean activations closer to zero and speeding up learning compared to ReLU.

Question 111: What problem does data augmentation specifically help address during transfer learning fine-tuning?

Catastrophic forgetting of source task
Gradient explosion in deep layers
Overfitting when the target dataset is small (Correct answer)
Slow convergence due to frozen layers

Correct answer: Overfitting when the target dataset is small

Data augmentation artificially expands the small target dataset, reducing overfitting during fine-tuning.

Question 112: 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 normalizes activations globally to prevent covariate shift
It averages across spatial dimensions to produce a single value per feature map, replacing dense layers (Correct answer)
It applies max pooling with a kernel equal to the full feature map size

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

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

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: What role does Batch Normalization play in CNN training?

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

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 115: 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 116: 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 reduces spatial dimensions like a strided convolution
It increases spatial dimensions, making it useful for upsampling (Correct answer)
It maintains identical spatial dimensions using same padding

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

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

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

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

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 119: 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?

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

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 120: A CNN model achieves 99% accuracy on the training dataset but only 75% on the validation dataset. Which phenomenon is occurring, and what is a common technique to address it?

Underfitting; increase model complexity by adding more layers.
Data leakage; remove features that are present in both training and validation sets.
Vanishing gradients; switch to a different activation function like Leaky ReLU.
Overfitting; apply Dropout to the fully connected layers. (Correct answer)

Correct answer: Overfitting; apply Dropout to the fully connected layers.

The significant performance gap between training and validation accuracy is a classic sign of overfitting, where the model has learned the training data too well, including its noise, and fails to generalize to new data. [24, 27] Dropout is a widely used regularization technique that randomly sets a fraction of neuron activations to zero during training, which helps prevent complex co-adaptations and improves generalization. [3, 18, 20]

Question 121: A pretrained CNN trained on ImageNet is being adapted for satellite imagery classification. Which transfer learning strategy is MOST appropriate?

Freeze all layers and only add a new softmax output
Use the model as-is with no retraining
Fine-tune all layers with a small learning rate (Correct answer)
Train from scratch ignoring ImageNet weights

Correct answer: Fine-tune all layers with a small learning rate

Satellite imagery differs significantly from ImageNet photos, so fine-tuning all layers (with a small LR) allows the model to adapt its features.

Question 122: What visualization technique overlays a heatmap on the input image to show which regions most influenced a CNN's prediction?

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

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 123: What distinguishes a ResNet 'basic block' from a 'bottleneck block'?

Basic blocks use 1x1 convolutions; bottleneck blocks use 3x3
Basic blocks include batch normalization; bottleneck blocks do not
Basic blocks are used in ResNet-50+; bottleneck blocks in ResNet-18/34
Basic blocks have two 3x3 convolutions; bottleneck blocks have a 1x1-3x3-1x1 sequence (Correct answer)

Correct answer: Basic blocks have two 3x3 convolutions; bottleneck blocks have a 1x1-3x3-1x1 sequence

Basic blocks (used in ResNet-18/34) stack two 3x3 convolutions, while bottleneck blocks (ResNet-50+) use a 1x1→3x3→1x1 pattern to reduce parameters.

Question 124: What does the 'random crop' augmentation technique do during CNN training?

Masks random channels
Reduces image resolution randomly
Crops a random sub-region of the image for training (Correct answer)
Removes random pixels from the image

Correct answer: Crops a random sub-region of the image for training

Random cropping extracts a randomly positioned sub-region of the input image, forcing the CNN to learn from different spatial locations.

Question 125: What distinguishes a depthwise separable convolution from a standard convolution?

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

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 126: Which of the following was a key innovation introduced in the AlexNet architecture that distinguished it from its predecessor, LeNet-5?

The introduction of the Inception module.
The use of average pooling for subsampling.
The use of the Rectified Linear Unit (ReLU) activation function. (Correct answer)
A shallow architecture with only 5 layers.

Correct answer: The use of the Rectified Linear Unit (ReLU) activation function.

While LeNet-5 used sigmoid or tanh activation functions, AlexNet was one of the first major architectures to popularize the use of the Rectified Linear Unit (ReLU). ReLU helped to significantly speed up the training process and mitigate the vanishing gradient problem, which was a major advantage for the deeper, 8-layer AlexNet architecture.

Question 127: Which data augmentation technique mixes two training images and their labels by linear interpolation to improve CNN generalization?

CutMix
Mixup (Correct answer)
CutOut
AutoAugment

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 128: In YOLOv3, how are predictions made at multiple scales?

By detecting at three different feature map scales using a FPN-like structure (Correct answer)
By ensembling outputs from three separate YOLO models
By applying three different anchor aspect ratios at a single scale
By running the network three separate times at different input resolutions

Correct answer: By detecting at three different feature map scales using a FPN-like structure

YOLOv3 predicts bounding boxes at three scales by extracting feature maps at different depths, similar to a feature pyramid network.

Question 129: Which layer type in a CNN is responsible for introducing non-linearity?

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

Correct answer: Activation layer

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

Question 130: What is group normalization and what problem does it solve compared to batch normalization?

It groups parameters and applies shared normalization across filter groups
It normalizes activations only within a group of related convolutional layers
It groups multiple batches together to compute more stable statistics
It normalizes groups of channels together, working well even with batch size of 1 (Correct answer)

Correct answer: It normalizes groups of channels together, working well even with batch size of 1

Group normalization divides channels into groups and normalizes within each group for each sample, making it effective for small batch sizes (like in object detection) where batch normalization degrades.

Question 131: How does increasing the batch size during CNN training affect regularization behavior?

Smaller batches produce noisier gradient estimates that act as implicit regularization (Correct answer)
Larger batches produce noisier gradients that act as regularization
Batch size has no effect on generalization
Larger batches always improve validation accuracy

Correct answer: Smaller batches produce noisier gradient estimates that act as implicit regularization

Smaller mini-batches introduce gradient noise that acts as implicit regularization, often leading to better generalization.

Question 132: What was the key innovation in the Faster R-CNN architecture that distinguished it from its predecessor, Fast R-CNN?

The replacement of the SVM classifier with a softmax layer for object classification.
The introduction of a Region Proposal Network (RPN) to generate object proposals within the main network. (Correct answer)
The implementation of RoI (Region of Interest) Pooling to handle inputs of different sizes.
The use of a deeper backbone network like VGG-16 for feature extraction.

Correct answer: The introduction of a Region Proposal Network (RPN) to generate object proposals within the main network.

The main bottleneck in Fast R-CNN was its reliance on an external, CPU-based algorithm like Selective Search to generate region proposals. The groundbreaking innovation of Faster R-CNN was the introduction of the Region Proposal Network (RPN), a fully convolutional network that is integrated into the main detection pipeline. The RPN shares convolutional features with the detection network, allowing it to generate high-quality region proposals almost for free, which made the entire object detection process significantly faster and trainable end-to-end.

Question 133: What is the role of Global Average Pooling (GAP) in modern transfer learning architectures?

Reduces each feature map to a single value, replacing flatten layers (Correct answer)
Performs data augmentation spatially
Computes attention weights across feature maps
Increases spatial resolution before the classifier

Correct answer: Reduces each feature map to a single value, replacing flatten layers

GAP averages each feature map spatially, reducing parameters and overfitting compared to large flatten-then-dense layers.

Question 134: In a typical Convolutional Neural Network (CNN) architecture, what is the primary function of the pooling layer?

To reduce the spatial dimensions of the feature maps. (Correct answer)
To perform classification based on extracted features.
To introduce non-linearity into the model.
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 135: Which of the following is a major advantage of using the ReLU (Rectified Linear Unit) activation function over the Sigmoid function in the hidden layers of a CNN?

It helps mitigate the vanishing gradient problem. (Correct answer)
It is more biologically plausible.
It outputs values in a normalized range of [0, 1].
It is a linear function, which simplifies computation.

Correct answer: It helps mitigate the vanishing gradient problem.

The ReLU function has a constant gradient of 1 for positive inputs, which helps to prevent the gradients from becoming extremely small during backpropagation in deep networks. The Sigmoid function, on the other hand, has gradients that are close to zero for very large or very small inputs, leading to the vanishing gradient problem.

Question 136: CutOut augmentation reduces CNN overfitting by which mechanism?

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

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

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

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

Same padding doubles the input size; valid padding halves it
Same padding uses larger filters; valid padding uses 1×1 filters
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

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

Scaling only depth to add more residual blocks uniformly
Jointly scaling network width, depth, and input resolution using a fixed ratio (Correct answer)
Scaling the learning rate and batch size together during training
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 140: 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?

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

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 141: In CNN regularization, what is 'stochastic depth' and how does it reduce overfitting in deep networks?

Applying variable Dropout rates to each layer
Gradually increasing the number of layers during training
Shrinking the depth of the network at test time
Randomly shortcircuiting (skipping) entire residual blocks during training (Correct answer)

Correct answer: Randomly shortcircuiting (skipping) entire residual blocks during training

Stochastic depth randomly bypasses entire residual layers during training, effectively training an ensemble of networks with different depths.

Question 142: What is the primary purpose of padding in a convolutional layer?

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

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

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

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

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

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

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

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 146: What is the total number of layers (depth) in the original GoogLeNet (Inception v1)?

22 (Correct answer)
12
34
17

Correct answer: 22

GoogLeNet is 22 layers deep (27 including pooling layers), which was very deep for its time when it won ILSVRC 2014.

Question 147: 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 148: 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 149: What is the fundamental difference between semantic segmentation and instance segmentation?

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

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 150: What are dilated (atrous) convolutions and why are they used in semantic segmentation?

Convolutions applied to dilated images
Convolutions with small kernels used to reduce computation
Convolutions used to separate color channels
Convolutions with gaps between kernel elements that expand the receptive field without increasing parameters or losing resolution (Correct answer)

Correct answer: Convolutions with gaps between kernel elements that expand the receptive field without increasing parameters or losing resolution

Dilated convolutions insert zeros between kernel weights (dilation rate > 1), allowing the filter to cover a larger area of the input without downsampling, which is critical for maintaining spatial resolution in segmentation.

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

These practice questions were written with AI tools. They are not real exam questions. Exam facts are checked against the official sources listed below. Not reviewed by a subject-matter expert. Editorial policy

Sources & references

Exam facts on this page (fees, format, dates, passing scores) are checked against these official sources; the date is our last check. Editorial policy

nncc-exam.org — Test delivery, retake policy, score release · verified Oct 2026
nncc-exam.org — Fees, exam format, passing score · verified Oct 2026