CNN Classic CNN Architectures 3 — Questions and Answers
Question 1: How does GoogLeNet reduce computational cost before applying 3x3 and 5x5 convolutions in the Inception module?
- By using strided convolutions
- By applying 1x1 convolutions as bottlenecks (Correct answer)
- By using average pooling before convolution
- By reducing the input image resolution
Correct answer: By applying 1x1 convolutions as bottlenecks
1x1 convolutions act as bottlenecks that reduce the number of input channels before more expensive 3x3 and 5x5 convolutions are applied.
Question 2: LeNet-5, designed for handwritten digit recognition, was trained on which dataset?
- ImageNet
- CIFAR-10
- MNIST (Correct answer)
- Pascal VOC
Correct answer: MNIST
LeNet-5 was developed by LeCun et al. to classify handwritten digits from the MNIST dataset, achieving state-of-the-art performance in the 1990s.
Question 3: Which activation function did AlexNet popularize as an alternative to sigmoid and tanh?
- Leaky ReLU
- ELU
- ReLU (Correct answer)
- Swish
Correct answer: ReLU
AlexNet popularized ReLU (Rectified Linear Unit), which trains faster than sigmoid/tanh and mitigates the vanishing gradient problem.
Question 4: What is the total number of layers (depth) in the original GoogLeNet (Inception v1)?
- 12
- 17
- 22 (Correct answer)
- 34
Correct answer: 22
GoogLeNet is 22 layers deep (27 including pooling layers), which was very deep for its time when it won ILSVRC 2014.
Question 5: What regularization technique unique to AlexNet was placed after the first two fully connected layers?
- L2 weight decay only
- Dropout with 0.5 probability (Correct answer)
- Batch normalization
- Label smoothing
Correct answer: Dropout with 0.5 probability
AlexNet applied dropout with p=0.5 after the first two fully connected layers, randomly zeroing activations during training to prevent co-adaptation.
Question 6: In a ResNet bottleneck block (used in ResNet-50+), what is the sequence of convolutional filter sizes?
- 3x3 → 3x3 → 3x3
- 1x1 → 3x3 → 1x1 (Correct answer)
- 1x1 → 1x1 → 3x3
- 3x3 → 1x1 → 3x3
Correct answer: 1x1 → 3x3 → 1x1
The ResNet bottleneck block uses 1x1 (reduce channels) → 3x3 (spatial convolution) → 1x1 (restore channels), reducing computation while maintaining representational power.
Question 7: What distinguishes a ResNet 'basic block' from a 'bottleneck block'?
- Basic blocks use 1x1 convolutions; bottleneck blocks use 3x3
- Basic blocks have two 3x3 convolutions; bottleneck blocks have a 1x1-3x3-1x1 sequence (Correct answer)
- Basic blocks include batch normalization; bottleneck blocks do not
- Basic blocks are used in ResNet-50+; bottleneck blocks in ResNet-18/34
Correct answer: Basic blocks have two 3x3 convolutions; bottleneck blocks have a 1x1-3x3-1x1 sequence
Basic blocks (used in ResNet-18/34) stack two 3x3 convolutions, while bottleneck blocks (ResNet-50+) use a 1x1→3x3→1x1 pattern to reduce parameters.
CNN Classic CNN Architectures 3
How does GoogLeNet reduce computational cost before applying 3x3 and 5x5 convolutions in the Inception module?