Module 06 Knowledge Check
5 questions. Pass mark 4 out of 5. Answer every question before checking the answer key below, then retry after reading the feedback.
1. A single convolution kernel…
- A. classifies the image
- B. asks one question of every patch
- C. reduces resolution
- D. normalises the input
2. Transfer learning works least well when…
- A. the source dataset was large
- B. your domain is far from the source domain
- C. you have many classes
- D. the images are small
3. When unfreezing a pretrained backbone you should…
- A. keep the same learning rate
- B. use a much lower learning rate
- C. increase the learning rate
- D. reinitialise the weights
4. Preprocessing constants should come from…
- A. memory or a tutorial
- B. the checkpoint's own documented transform
- C. your data's statistics
- D. the framework default
5. Before training, you should compute…
- A. the learning rate schedule
- B. the majority-class baseline
- C. the model size
- D. the number of epochs
---
Answer Key and Explanations
Check these only after attempting every question.
1. B - asks one question of every patch
A layer with 64 kernels asks 64 questions at every position; the next layer asks questions about the answers.
2. B - your domain is far from the source domain
Early layers still help; late layers encode whole-object concepts with no counterpart in, say, a radiograph.
3. B - use a much lower learning rate
The original rate pushes large gradients through weights that took enormous compute to learn.
4. B - the checkpoint's own documented transform
They are properties of the checkpoint, and a mismatch degrades silently with no error.
5. B - the majority-class baseline
Three architectures at 91, 92 and 94% look like progress until you notice the constant predictor scores 94.
Practical Check
Apply this module to your own work: complete the module activity for *CNN Intuition and Transfer Learning*, then write one sentence naming what your result shows and one naming what it does not.
Strong Answer Pattern
A strong answer names the task, the evidence used, the check performed, and the remaining limitation. It avoids "proved", "guaranteed", or "always" unless the evidence genuinely supports it.