Bottleneck Architectures Quiz

Answers and explanations appear as you take the quiz.

1. In a ResNet bottleneck block, what is the role of the first 1x1 convolution?
   • To reduce the number of channels before the 3x3 convolution
   • To increase spatial resolution
   • To apply the final classification
   • To add noise for regularization
2. Why does a bottleneck make deep networks cheaper to compute?
   • It removes all activations
   • It performs expensive operations on fewer channels
   • It uses larger kernels everywhere
   • It skips backpropagation
3. In an autoencoder, the bottleneck layer primarily determines:
   • The optimizer choice
   • The image file format
   • How much the input is compressed
   • The batch size
4. MobileNetV2 uses which variant of the bottleneck idea?
   • Pure pooling bottleneck
   • Recurrent bottleneck
   • No bottleneck at all
   • Inverted residual bottleneck (expand then contract)
5. How does LoRA apply bottleneck principles to large language models?
   • By inserting a trainable low-rank (narrow) bottleneck into weight updates
   • By deleting attention layers
   • By doubling all weights
   • By quantizing to 1 bit
6. A typical ResNet bottleneck block follows which channel pattern?
   • Wide, wide, wide
   • Reduce, process, restore (e.g., 256 to 64 to 256)
   • Narrow, wide, narrow
   • Random per layer
7. A ResNet block reduces 256 channels to 64 before its 3×3 convolution. Ignoring the 1×1 layers, what is the 3×3 layer’s approximate multiply-add cost relative to 256 channels?
   • About one-quarter
   • About one-half
   • About one-sixteenth
   • About four times as much
8. An autoencoder uses a narrow latent code. What constraint does this impose on reconstruction?
   • The encoder must preserve every input pixel exactly
   • The network can no longer use learned weights
   • The decoder must expand the latent code before encoding
   • The decoder must reconstruct the input from a compressed representation