Module 10 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. Two vectors point the same direction, one twice as long. What do cosine and Euclidean say?
- A. Both say identical
- B. Cosine says identical; Euclidean says they are apart
- C. Both say apart
- D. Cosine errors
2. Reconstruction error collapses at bottleneck width 2 and barely improves at 4. What does that indicate?
- A. The autoencoder is broken
- B. The data's intrinsic dimensionality is about 2
- C. The learning rate is wrong
- D. More epochs are needed
3. Low reconstruction error proves the codes are useful for your downstream task.
- A. True
- B. False — a model can reconstruct perfectly while encoding nothing the task needs
- C. True only for images
- D. True only above the intrinsic dimension
4. A classifier recovers a sensitive attribute from your embeddings at 0.95 against a 0.55 base rate. What follows?
- A. Nothing — you did not use it as a feature
- B. The attribute is encoded and is influencing retrieval and ranking
- C. The probe is overfitting
- D. Normalise the vectors
5. Why record the model version alongside cached embeddings?
- A. For licensing
- B. Vectors from two checkpoints are not comparable, even for the same input
- C. To save disk space
- D. It is not necessary
---
Answer Key and Explanations
Check these only after attempting every question.
1. B — Cosine says identical; Euclidean says they are apart
Cosine ignores length, Euclidean does not. Normalising to unit length makes them agree, which is why embedding pipelines normalise.
2. B — The data's intrinsic dimensionality is about 2
Below the true dimensionality information is destroyed. Above it, the extra capacity has nothing left to encode.
3. B — False — a model can reconstruct perfectly while encoding nothing the task needs
Reconstruction quality and downstream usefulness are separate questions. Test the codes on the task directly.
4. B — The attribute is encoded and is influencing retrieval and ranking
Correlated features reconstruct an attribute you never supplied. The probe is the only way to find out.
5. B — Vectors from two checkpoints are not comparable, even for the same input
Mixing encodings degrades retrieval in a way that is very hard to trace back to its cause.
Practical Check
Apply this module to your own work: complete the module activity for *Representation Learning and Embedding Workflows*, 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.