ML Chapter 08

Why are deeper networks more effective in deep learning?

• They have fewer weights
• They skip training
• They can learn hierarchical features
• They use fewer neurons

What is a main challenge when using deep networks?

• Fast overfitting
• Difficulty in capturing low-level patterns
• Vanishing gradients
• Lack of enough parameters

What technique helps reduce vanishing gradients?

• Pooling
• Dropout
• Skip connections (ResNet)
• Softmax

What is the purpose of data augmentation?

• Reduce training time
• Increase the number of training samples artificially
• Make the model smaller
• Apply dropout to images

Which of the following is a data augmentation method?

• Freeze batchnorm
• Max pooling
• Image rotation
• Weight decay

What does a deep network learn in early layers?

• Object categories
• Full image context
• Edges and simple patterns
• Game strategies

Which model marked a breakthrough in 2012 using ImageNet?

• LeNet
• VGG
• AlexNet
• ResNet

What is the characteristic of VGG networks?

• Filters of multiple sizes
• 3x3 filters and consistent blocks
• Very shallow layers
• Uses reinforcement learning

What did GoogLeNet introduce to deep learning?

• Residual learning
• 1-layer networks
• Inception modules with multiple filter sizes
• Batch normalization

What makes ResNet special compared to earlier CNNs?

• Uses fewer parameters
• Combines LSTM with CNNs
• Introduces skip connections to allow very deep networks
• Is the shallowest model