/
Deep Learning Concepts
Save to my account
Sign up
Deep Learning Concepts
Deep Learning Concepts
Study
1
Question
What does a convolutional layer use filters to perform?
Answer
Convolution operations scanning input for features.
2
Question
What is the purpose of pooling layers in CNNs?
Answer
Downsample feature maps and introduce spatial invariance.
3
Question
How is the output size of a convolution computed?
Answer
O = I - F + Pstart + Pend + 1.
4
Question
What parameter determines the movement of the window in convolution?
Answer
Stride (S).
5
Question
What is a key feature of the ReLU activation function?
Answer
Outputs max(0, z), adding non-linearity.
6
Question
How does max pooling differ from average pooling?
Answer
Max pooling selects maximum values; average pooling computes mean.
7
Question
What role do zero-padding modes like 'valid' and 'same' play in CNNs?
Answer
Control feature map size and feature detection.
8
Question
What is a fully connected (FC) layer's main characteristic?
Answer
Each input connects to all neurons.
9
Question
What is the main purpose of the softmax function in neural networks?
Answer
Convert scores into probability distributions.
10
Question
What does the receptive field in a CNN indicate?
Answer
Input area influencing each activation map pixel.
11
Question
What is the primary purpose of object detection algorithms like YOLO and R-CNN?
Answer
Classify objects and locate them with bounding boxes.
12
Question
How does IoU measure the accuracy of predicted bounding boxes?
Answer
Computes intersection over union of boxes.
13
Question
What is the primary difference between face verification and face recognition?
Answer
Verification is one-to-one; recognition is one-to-many.
14
Question
What is the purpose of the Gram matrix in neural style transfer?
Answer
Quantifies channel correlation for style.
15
Question
What do content and style loss functions optimize in neural style transfer?
Answer
Content: similarity with original image; Style: similarity with style image.
16
Question
What are a Generative Adversarial Network's main components?
Answer
Generator and discriminator.
17
Question
In RNNs, what problem does the vanishing gradient cause?
Answer
Difficulty capturing long-term dependencies.
18
Question
What is the purpose of the forget gate in LSTM?
Answer
To erase or retain cell state information.
19
Question
How does the skip-gram model in word2vec learn embeddings?
Answer
By predicting context words given a target word.
20
Question
What is perplexity in language modeling?
Answer
Inverse probability of the dataset; lower is better.
21
Question
What does cosine similarity measure between two word vectors?
Answer
Cosine of the angle between them.
22
Question
What is the main goal of attention mechanisms in models?
Answer
Focus on important input parts.
23
Question
What does the hyperparameter 'beam width' in beam search control?
Answer
Trade-off between search quality and computational cost.
24
Question
How does dropout regularization help neural networks?
Answer
Prevents overfitting by randomly deactivating neurons.
25
Question
What is the purpose of data augmentation?
Answer
Increase training data variability artificially.
26
Question
How does batch normalization improve training?
Answer
Normalizes layer inputs to stabilize learning.
27
Question
What is the primary purpose of cross-entropy loss?
Answer
Measure divergence between predicted and true distributions.
28
Question
How does backpropagation update weights?
Answer
Computes gradients via chain rule and adjusts weights accordingly.
29
Question
What role does batch normalization play in training?
Answer
Reduces dependence on initialization and allows higher learning rates.
30
Question
What is Xavier initialization designed to do?
Answer
Set initial weights considering layer characteristics for stability.