chatGPT questions exam2

Question 1

What does Stochastic Gradient Descent (SGD) optimize in machine learning models?

Answer 1

SGD is used to minimize the loss function of a model.

Question 2

How does SGD differ from traditional gradient descent?

Answer 2

Unlike traditional gradient descent that uses the entire dataset for updates, SGD updates model parameters using only a single sample or a small batch of samples.

Question 3

What are the key benefits and drawbacks of using SGD?

Answer 3

SGD is more computationally efficient for large datasets and can help escape local minima, but it may lead to slower and less stable convergence.

Question 4

What does the FROC curve evaluate in medical image analysis?

Answer 4

the FROC curve evaluates detection and localization performance, plotting sensitivity against the average number of false positives per image.

Question 5

Why can accuracy be a misleading performance metric?

Answer 5

Accuracy can be misleading in cases of class imbalance, where it doesn’t reflect the model’s performance on the minority class

Question 6

What was the major limitation of the original R-CNN in object detection?

Answer 6

The original R-CNN was slow due to its reliance on selective search for generating region proposals.

Question 7

How did Fast R-CNN improve over R-CNN?

Answer 7

Fast R-CNN improved efficiency by sharing computations across region proposals.

Question 8

What innovation did Faster R-CNN introduce to object detection?

Answer 8

Faster R-CNN introduced a Region Proposal Network (RPN), allowing for end-to-end training and faster proposal generation.

Question 9

How does setting a high threshold for predictions affect sensitivity and specificity?

Answer 9

A high threshold generally increases specificity but may decrease sensitivity by rejecting true positives with lower confidence.

Question 10

What approach do convolutional networks use for segmentation tasks?

Answer 10

They utilize architectures with downsampling and upsampling layers, and sometimes strided or dilated convolutions, to produce a pixel-wise segmentation map.

Question 11

How is the number of trainable parameters in a convolutional layer calculated?

Answer 11

The formula is (filter height × filter width × input channels + 1) × number of filters, where “+1” accounts for the bias term.

Question 12

Why is the learning rate considered a critical hyperparameter in neural network training?

Answer 12

t influences the training dynamics, where too high a rate may cause divergence and too low a rate results in slow convergence.

Question 13

What does dropout do in neural networks?

Answer 13

Dropout prevents overfitting by randomly dropping units and their connections during training, simulating training multiple networks in parallel.

Question 14

What are the advantages of ReLU over sigmoid functions?

Answer 14

ReLU mitigates the vanishing gradient problem, accelerates convergence, and maintains gradient flow for positive inputs.

Question 15

Why is feature normalization important before training a machine learning model?

Answer 15

it scales features to have zero mean and unit variance, ensuring consistent scaling and preventing information leakage from the test set.

Question 16

How do residual networks (ResNets) facilitate the training of deep models?

Answer 16

resNets use skip connections to learn residual mappings, alleviating the vanishing gradient problem and enabling deeper architectures.

Question 17

What does the receptive field refer to in convolutional networks?

Answer 17

It refers to the size of the input area that influences the network’s output, determined by the cumulative effect of convolution and pooling operations.

Question 18

How does the softmax function work in neural networks?

Answer 18

It converts raw output scores into probabilities by taking the exponential of each output and normalizing these values by the sum of all exponentials.

Question 19

What does a false negative (FN) indicate in detection tasks?

Answer 19

A FN occurs when an object is present but not detected by the classifier, which is critical in applications where missing detections can have serious consequences.

Question 20

What is a dilated (atrous) convolution and its purpose?

Answer 20

It increases the receptive field without increasing the number of weights by inserting spaces between kernel elements, allowing for broader spatial aggregation.

Question 21

How does U-Net architecture specialize for medical image segmentation?

Answer 21

U-Net uses a symmetric architecture with downsampling and upsampling paths, enabling precise localization and context integration without needing a pre-defined weight map.

Question 22

How is specificity calculated in a model’s performance evaluation?

Answer 22

Specificity is calculated as the number of true negatives divided by the sum of true negatives and false positives.

Question 23

What determines the output size of a valid convolution operation?

Answer 23

The output size is determined by (W − F + 1) × (H − F + 1), where W and H are the width and height of the image, and F is the filter size.

Question 24

What makes the YOLO object detection framework unique?

Answer 24

YOLO processes the entire image in a single evaluation and makes predictions for each grid cell, combining bounding box predictions and class probabilities.

Question 25

How does backpropagation work in training neural networks?

Answer 25

It computes gradients of the loss with respect to network parameters by applying the chain rule, allowing for efficient parameter updates.

Question 26

What are histogram-based features in image processing?

Answer 26

These features, like mean, variance, skewness, and kurtosis, describe the distribution of pixel intensities and are invariant to/do not consider spatial relationship and correlation between pixels (identical histograms can belong to different textures)

Question 27

how does the shift-and-stitch algorithm benefit fully convolutional networks?

Answer 27

It enables the generation of fine-grained outputs by reconstructing the original resolution through multiple shifted input passes.

Question 28

What improvement does Fast R-CNN offer over the original R-CNN?

Answer 28

Fast R-CNN processes the entire image at once and uses RoI pooling to extract features, improving efficiency and speed

Question 29

How does histogram matching adjust an image’s brightness and contrast?

Answer 29

It modifies an image so that its histogram matches that of a reference image, using cumulative distribution functions for mapping pixel values.

Question 30

What are the implications of not using pooling layers in CNNs?

Answer 30

Omitting pooling layers requires convolutional layers with strides for downsampling, potentially leading to higher memory and computational costs.

Question 31

How does backpropagation facilitate learning in CNNs?

Answer 31

It calculates gradients (using chain rule) for all network parameters by propagating errors backward through the network, enabling parameter updates for learning.

Question 32

What are the steps in mini-batch gradient descent training?

Answer 32

The steps are selecting a mini-batch, performing a forward pass, computing loss, backpropagating to compute gradients, and updating weights.

Question 33

How does the choice of pooling size and stride affect max-pooling operations?

Answer 33

Incompatible pooling sizes and strides may lead to suboptimal coverage and downsampling uniformity in the feature map.

Question 34

What strategy is used in transfer learning with small datasets?

Answer 34

Freeze most of the pre-trained network’s parameters and fine-tune the last few layers to adapt the features to the new task.

Question 35

How does dilated convolution affect the receptive field in CNNs?

Answer 35

dilated convolution expands the receptive field without increasing the number of weights, allowing for broader spatial information aggregation.

Question 36

What is the purpose of the softmax function in neural networks?

Answer 36

it converts logits into probabilities, facilitating multi-class classification by ensuring output values sum to 1.

Question 37

What does the term “dying ReLU” problem refer to?

Answer 37

it refers to the issue where neurons stop learning due to always outputting zero, which can stem from improper initialization or gradient descent issues. - this can be solved with leakyrelu

Question 38

How is the number of trainable parameters in a dense layer calculated?

Answer 38

It is calculated as (number of inputs + 1) × number of outputs, accounting for each input connection and bias term.

Question 39

What does zero padding do in convolutional layers?

Answer 39

Zero padding allows control over output dimensions, maintaining spatial sizes or adjusting them for specific convolutional effects.

Question 40

How is the “best model” identified during neural network training?

Answer 40

The best model is often the one with the lowest validation loss, indicating effective learning and generalization capabilities without overfitting.