Neural Network Deep Learning Architectures 2 β Questions and Answers
Question 1: What problem does batch normalization primarily address in deep networks?
- Overfitting on training data
- Internal covariate shift between layers (Correct answer)
- Vanishing gradients only
- Slow inference speed
Correct answer: Internal covariate shift between layers
Batch normalization reduces internal covariate shift by normalizing layer inputs, stabilizing and accelerating training.
Question 2: Which architecture introduced the concept of inception modules combining multiple filter sizes in parallel?
- VGG
- AlexNet
- GoogLeNet (Inception) (Correct answer)
- ResNet
Correct answer: GoogLeNet (Inception)
GoogLeNet introduced inception modules that apply 1Γ1, 3Γ3, and 5Γ5 convolutions in parallel to capture multi-scale features.
Question 3: What is the purpose of the bottleneck layer in an autoencoder?
- To speed up training
- To force the network to learn a compressed representation (Correct answer)
- To apply attention mechanisms
- To perform classification
Correct answer: To force the network to learn a compressed representation
The bottleneck layer constrains the network to learn a compressed, meaningful representation of the input by limiting the number of neurons.
Question 4: In a U-Net architecture, what is the purpose of the skip connections between encoder and decoder?
- To reduce computational cost
- To preserve fine-grained spatial information during reconstruction (Correct answer)
- To apply dropout during decoding
- To normalize batch statistics
Correct answer: To preserve fine-grained spatial information during reconstruction
Skip connections in U-Net pass high-resolution feature maps from the encoder directly to the decoder, preserving spatial detail for precise segmentation.
Question 5: What distinguishes a capsule network from a standard CNN?
- It uses recurrent connections
- It encodes pose information as vectors rather than scalar activations (Correct answer)
- It applies self-attention to feature maps
- It uses no pooling operations at all
Correct answer: It encodes pose information as vectors rather than scalar activations
Capsule networks represent features as vectors (capsules) that encode both the presence and spatial relationships/pose of entities, unlike scalar activations in CNNs.
Question 6: Which regularization technique randomly drops entire feature map channels during training in CNNs?
- Weight decay
- Spatial dropout (Correct answer)
- L1 regularization
- Gradient clipping
Correct answer: Spatial dropout
Spatial dropout drops entire feature map channels (2D feature maps) rather than individual neurons, which is more effective for convolutional layers.
Neural Network Deep Learning Architectures 2
What problem does batch normalization primarily address in deep networks?