Residual Networks Quiz

Answers and explanations appear as you take the quiz.

1. What problem did residual connections specifically solve?
   • Deeper plain networks training worse than shallower ones (degradation)
   • Images being too low resolution
   • A shortage of labeled training data
   • GPUs running out of memory
2. What does a basic residual block compute when input and output dimensions match?
   • Multiplies the input by itself
   • Adds the block’s learned residual function to its input, y = F(x) + x
   • Drops the input and returns zero
   • Sorts input channels alphabetically
3. Why do skip connections help gradients during training?
   • They eliminate the need to train weights
   • They force every layer to output zero
   • They provide an additional direct path for gradient flow around the learned transformation
   • They make deep networks immune to all optimization problems
4. How deep was the deepest ResNet evaluated in the original ImageNet paper?
   • 18 layers
   • 34 layers
   • 50 layers
   • 152 layers
5. If a residual block's layers learn F(x) = 0, what does the block do?
   • It effectively passes the input through unchanged (identity)
   • It outputs zero and breaks the network
   • It doubles the input
   • It randomizes the output
6. What residual connection does the Transformer paper describe around attention and feed-forward sublayers?
   • Replace the sublayer input with a class label
   • Add each sublayer’s output to the sublayer input through a residual connection
   • Average every token globally at the output only
   • Remove the residual stream after self-attention
7. When a residual block changes channel dimensions, what can a projection shortcut do?
   • Convert the input into a text sequence
   • Remove the residual branch
   • Map the shortcut input to the dimensions required for addition
   • Guarantee that every gradient has unit magnitude
8. What sequence of convolutions defines the original ResNet bottleneck block?
   • Three 3×3 convolutions only
   • A 3×3 convolution followed by two fully connected layers
   • A pooling layer followed by two transposed convolutions
   • A 1×1, then 3×3, then 1×1 convolution