chatGPT questions exam1

Question 1

What is Batch Normalization and its purpose in neural networks?

Answer 1

Batch Normalization (BN) is a technique that normalizes the outputs of a previous activation layer by subtracting the batch mean and dividing by the batch standard deviation, including learnable parameters for scale and shift to maintain the network’s representational capacity. It improves speed, performance, and stability.

Question 2

What does a convolution operation involve in CNNs?

Answer 2

In CNNs, convolution involves sliding a filter over the input feature map and computing the dot product at every position to extract features

Question 3

Why is padding used in convolutional neural networks?

Answer 3

Padding is used to adjust the size of the feature maps so that the output feature map can maintain the same size as the input, allowing for the application of the filter at the border of the input.

Question 4

What is Connected Component Analysis used for in image processing?

Answer 4

Connected Component Analysis labels and groups pixels into components based on pixel connectivity for segmentation tasks, aiding in object detection and segmentation.

Question 5

What is Categorical Cross-Entropy and its role in classification problems?

Answer 5

Categorical Cross-Entropy is a loss function measuring the dissimilarity between true distributions and predicted probabilities, penalizing incorrect predictions in multi-class classification problems.

Question 6

What does the Max-Pooling operation do in CNNs?

Answer 6

Max-Pooling is a down-sampling operation that selects the maximum value within a window, reducing dimensionality and highlighting the most activated features.

Question 7

What does the ROC curve represent in binary classification systems?

Answer 7

The ROC curve plots the true positive rate against the false positive rate at various threshold settings, showing the trade-off between sensitivity and specificity.

Question 8

How does the ReLU activation function mitigate the vanishing gradient problem?

Answer 8

ReLU (Rectified Linear Unit) mitigates the vanishing gradient problem with its linear, non-saturating form, preventing gradients from becoming very small during training.

Question 9

What is the purpose of hard-negative mining in machine learning?

Answer 9

Hard-negative mining focuses on selecting difficult-to-classify negative samples, improving model performance by ensuring it learns from the most challenging examples.

Question 10

How does dropout help in preventing overfitting in neural networks?

Answer 10

Dropout prevents overfitting by randomly dropping units and their connections during training, encouraging the network to learn robust features.

Question 11

List one strategy to combat overfitting in neural networks.

Answer 11

Implementing L2 regularization adds a penalty on large weights to the loss function, helping to prevent the model from overfitting.

Question 12

What information does the histogram of an image provide?

Answer 12

An image histogram displays the frequency of pixel intensities, offering insights into the contrast, brightness, and intensity distribution of the image.

Question 13

How are trainable parameters calculated in a convolutional layer?

Answer 13

In a convolutional layer, the number of parameters is calculated as (filter width × filter height × input depth + 1) × number of filters, where “+1” accounts for the bias term.

Question 14

What characterizes a valid convolution operation?

Answer 14

Valid convolution ensures the filter fits entirely within the input image bounds, leading to a smaller output size.

Question 15

What does sensitivity measure in a binary classification system?

Answer 15

Sensitivity, or true positive rate, measures the proportion of actual positives correctly identified by the classifier.

Question 16

What is binary morphology used for in image processing?

Answer 16

Binary morphology uses structuring elements to manipulate shapes in binary images for purposes like noise removal, shape analysis, and segmentation.

Question 17

What is the key characteristic of mini-batch Stochastic Gradient Descent?

Answer 17

Mini-batch SGD updates model parameters for each small batch of the training set, balancing computational efficiency and update frequency.

Question 18

Name a factor that has contributed to the progress in deep learning.

Answer 18

The availability of powerful computational resources like GPUs, which accelerate the training of deep neural networks and activation functions.

Question 19

What is the U-Net architecture designed for, and how does it handle spatial dimensions?

Answer 19

• The U-Net architecture is designed for biomedical image segmentation. It features a contracting path and an expanding path with upsampling layers and skip connections to preserve spatial dimensions for precise localization.

Question 20

What is the structure of a neural network with three layers?

Answer 20

A neural network with three layers consists of an input layer, two hidden layers, and an output layer, where the input layer receives data, the hidden layers process it, and the output layer produces the final prediction.

Question 21

What is the K-means algorithm used for?

Answer 21

K-means is an unsupervised learning algorithm used for clustering, aiming to partition observations into clusters with the nearest means.

Question 22

What is the purpose of histogram matching in image processing?

Answer 22

Histogram matching adjusts the brightness and contrast of an image to match the histogram of a target image, used for normalization and image stitching.

Question 23

How do CNNs differ from traditional neural networks in handling data?

Answer 23

CNNs, specialized for analyzing visual imagery, use weight sharing and convolutional layers to efficiently extract local features from image data.

Question 24

How did Faster R-CNN improve upon R-CNN and Fast R-CNN?

Answer 24

Faster R-CNN integrated a Region Proposal Network (RPN) for nearly cost-free region proposals, enabling end-to-end training and improving object detection efficiency.

Question 25

What is a common sequence of operations in designing CNN architectures?

Answer 25

A common sequence involves a convolutional layer followed by an activation function like ReLU, then batch normalization for stability, and dropout for regularization.

Question 26

How can the receptive field be increased in convolutional networks?

Answer 26

The receptive field can be increased using dilated (or atrous) convolution, which involves inserting spaces between kernel elements to capture larger input areas without increasing parameters.

Question 27

What does the convolution operation achieve in CNNs?

Answer 27

The convolution operation extracts features like edges and textures by sliding a filter over the input and computing the sum of element-wise multiplications.

Question 28

How does dilation affect the output size calculation in CNNs?

Answer 28

Dilation increases the receptive field without significantly reducing spatial size, affecting the output size calculation by considering filter size, stride, padding, and dilation factor.

Question 29

What role do anchors play in object detection models like YOLO and Faster R-CNN?

Answer 29

Anchors facilitate the detection of various object sizes and aspect ratios by allowing multiple classes and bounding boxes predictions for each anchor, improving detection accuracy.

Question 29

Why are GPU memory requirements higher during training than testing in neural networks?

Answer 29

During training, networks need to store variables for the forward pass and gradients for the backward pass, while testing only requires forward pass calculations, reducing memory requirements.

Question 30

How do convolutional filters differ from neurons in neural networks?

Answer 30

Convolutional filters share weights across all positions to detect features efficiently, whereas neurons in fully connected layers do not share weights, leading to higher parameter counts.

Question 31

What is Leave-One-Out Cross-Validation and when is it useful?

Answer 31

LOOCV is a method where the model is trained N times on N-1 samples and tested on the remaining sample, useful for small datasets to maximize training data use.

Question 32

What is the function of a 2D Gaussian filter in image processing?

Answer 32

A 2D Gaussian filter smooths or blurs images by averaging pixels under a Gaussian-shaped kernel, reducing noise and detail.

Question 33

How does dropout function as a regularization technique?

Answer 33

Dropout randomly drops units during training, simulating a sparse activation and forcing the network to learn robust features, preventing overfitting.

Question 34

What are common causes of overfitting in neural networks?

Answer 34

Overfitting often occurs when a high-capacity model is trained on a small dataset, leading to poor generalization to new data.

Question 35

How are trainable parameters calculated between two dense layers?

Answer 35

The number of trainable parameters is calculated as (input size + 1) × output size, accounting for the bias term in the current layer.

Question 36

**How can a feature map size be reduced using a convolutional layer?

Answer 36

• To reduce the size of a feature map, use 1x1 convolutions with a specific number of filters. For example, to reduce a feature map of size 32x32x64 to 32x32x16, use 16 filters of size 1x1x64, mixing input channels without altering spatial dimensions.

Question 37

How is the receptive field calculated in convolutional networks with multiple layers?

Answer 37

The receptive field is calculated considering the size of the filters and the stride with which they are applied. It increases with each convolutional layer, allowing for a larger contextual understanding for predictions.

Question 38

How do sensitivity and specificity relate to the acceptance of model predictions?

Answer 38

Accepting all predictions with a score greater than 0 typically results in high sensitivity by capturing most true positives, but may also lead to low specificity due to an increase in false positives.

Question 39

What are the main advantages of convolutional layers in neural networks?

Answer 39

Convolutional layers efficiently use parameters through weight sharing and are adept at extracting local features from image data, making them particularly effective for computer vision tasks.