Squeeze-and-Excitation Networks Quiz

Answers and explanations appear as you take the quiz.

1. What does the 'squeeze' step in an SE block do?
   • Uses global average pooling to summarize each channel into one value
   • Crops the image to a smaller size
   • Removes redundant channels permanently
   • Applies a softmax over spatial locations
2. What does the 'excitation' step output?
   • A new set of convolution filters
   • A per-channel weight between 0 and 1
   • A bounding box
   • A segmentation mask
3. How are the excitation weights applied to the original feature maps?
   • Added element-wise
   • Concatenated as new channels
   • Multiplied channel-wise
   • Used to mask out spatial regions
4. Why does the excitation use a bottleneck with a reduction ratio (often 16)?
   • To increase the number of output classes
   • To remove the need for convolutions
   • To make the network shallower
   • To keep the added parameters and compute small
5. What competition did SENet win in 2017?
   • The ImageNet ILSVRC classification challenge
   • The COCO detection challenge
   • The Kaggle Titanic competition
   • The GLUE benchmark
6. Which later architectures embed SE blocks for efficiency?
   • BERT and GPT
   • EfficientNet and MobileNetV3
   • Word2Vec and GloVe
   • WaveNet and Tacotron
7. In the excitation step, which activations connect the two fully connected layers in the guide’s SE block?
   • Softmax followed by ReLU
   • Tanh followed by max pooling
   • ReLU followed by sigmoid after the second projection
   • GELU followed by layer normalization
8. An SE block has C = 256 channels and reduction ratio r = 16. What is the bottleneck width?
   • 4 units
   • 32 units
   • 256 units
   • 16 units