Special Applications: Face Recognition & Neural Style Transfer:Convolutional Neural Networks(Deep Learning Specialization)Answers:2025

Question 1

Which of the following do you agree with?

• ❌ Face recognition requires comparing pictures against one person’s face.

• ✅ Face recognition requires K comparisons of a person’s face.

• ❌ Face verification requires K comparisons of a person’s face.

Explanation:
Face recognition (identifying who the person is among $K$ known identities) generally requires comparing the probe face against the $K$ stored identities (or computing similarity to each of $K$ templates). Face verification (is this person X?) requires just a single comparison to the claimed identity, not $K$.

---

Question 2

Workgroup membership detection — which do you agree with?

• ✅ It will be more efficient to learn a function d(img1,img2)d(\text{img}_1,\text{img}_2)d(img1​,img2​) for this task.

• ❌ It is best to build a CNN with a softmax output with as many outputs as members of the group.

• ❌ This can’t be considered a one-shot learning task since there might be many members in the workgroup.

• ✅ This can be considered a one-shot learning task.

Explanation:
A learned similarity/distance function $d(\cdot ,\cdot )$ (or embedding + nearest-neighbor) is flexible: you can add/remove members without retraining the classifier head. This is exactly the one-shot / few-shot setup: you compare a probe to one (or a few) examples per person rather than training a fixed softmax over all members.

---

Question 3

To train with triplet loss, must you collect pictures only from current team members? True/False?

• ✅ False

• ❌ True

Explanation:
Triplet-loss training benefits from many distinct identities to teach the embedding to separate different faces. You do not need to restrict triplet training to only current workgroup members — using many other identities (external data) typically improves embedding quality and generalization.

---

Question 4

Triplet loss max⁡(∥f(A)−f(P)∥2−∥f(A)−f(N)∥2+α,0)\max(\|f(A)-f(P)\|^2 – \|f(A)-f(N)\|^2 + \alpha, 0)max(∥f(A)−f(P)∥2−∥f(A)−f(N)∥2+α,0) is larger in which case?

• ❌ When the encoding of A is closer to the encoding of P than to the encoding of N.

• ✅ When the encoding of A is closer to the encoding of N than to the encoding of P.

• ❌ When $A=P$ and $A=N$.

Explanation:
The loss grows when the positive is farther from A than the negative: i.e., ∥f(A)−f(P)∥2\|f(A)-f(P)\|^2∥f(A)−f(P)∥2 is large relative to ∥f(A)−f(N)∥2\|f(A)-f(N)\|^2∥f(A)−f(N)∥2. Intuitively, loss is large when A is closer to N than to P (violating the desired inequality). If $A=P$ and $A=N$ (identical points), the expression reduces to $\alpha$ (if both distances are 0), but the typical failure case is when negative is too close.

---

Question 5

Siamese network architecture — which do you agree with most?

• ✅ The upper and lower neural networks depicted have exactly the same parameters, but the outputs are computed independently for each image.

• ❌ This depicts two different neural networks with different architectures.

• ❌ The two networks have the same architecture, but they might have different parameters.

• ❌ The two images are combined in a single volume and pass through a single neural network.

Explanation:
A Siamese network uses shared weights — the same network (same parameters) applied to each input separately to produce embeddings; outputs are computed independently and then compared (e.g., distance).

---

Question 6

You’re more likely to find a unit responding strongly to cats in layer 4 than layer 1. True/False?

• ✅ True

• ❌ False

Explanation:
Lower layers (layer 1) learn low-level features (edges, colors); deeper layers (layer 4) capture higher-level, semantically meaningful features (parts or whole-object detectors), so you’re more likely to find a neuron selective for cats in deeper layers.

---

Question 7

Neural style transfer: which loss terms are used? (choose all that apply)

• ✅ JstyleJ_{\text{style}}Jstyle​ that compares $S$ and $G$.

• ❌ JcorrJ_{\text{corr}}Jcorr​ that compares $C$ and $S$.

• ❌ $T$ that calculates triplet loss between $S,G,C$.

• ✅ JcontentJ_{\text{content}}Jcontent​ that compares $C$ and $G$.

Explanation:
Neural style transfer optimizes a generated image $G$ to minimize: content loss Jcontent(G,C)J_{content}(G,C)Jcontent​(G,C) (match high-level content features of $C$) and style loss Jstyle(G,S)J_{style}(G,S)Jstyle​(G,S) (match Gram-matrix/style statistics of $S$). Triplet/correlation terms are not part of standard NST.

---

Question 8

Content loss Jcont(G,C)=∥a[l](C)−a[l](G)∥2J_{cont}(G,C) = \|a^{[l]}(C) – a^{[l]}(G)\|^2Jcont​(G,C)=∥a[l](C)−a[l](G)∥2. We choose $l$ to be a very high value to use the more abstract activation. True/False?

• ✅ True

• ❌ False

Explanation:
For content similarity you pick a deeper layer $l$ (higher-level activations) because they capture abstract content/structure rather than low-level texture; hence $l$ is chosen among deeper layers.

---

Question 9

In neural style transfer, what is updated each iteration?

• ❌ The pixel values of the content image $C$

• ❌ The neural network parameters

• ✅ The pixel values of the generated image $G$

• ❌ The regularization parameters

Explanation:
In NST the pretrained network parameters are fixed; optimization is performed over the image pixels of $G$ to minimize combined content+style loss.

---

Question 10

3D conv: input $32\times 32\times 32\times 3$, 16 filters of size $4\times 4\times 4$, zero padding and stride 1. Output size?

• ❌ $31\times 31\times 31\times 16$

• ✅ $29\times 29\times 29\times 16$

• ❌ $29\times 29\times 29\times 13$

• ❌ $29\times 29\times 29\times 3$

Explanation:
Assuming zero padding = 0 (no padding), spatial output dims = $(32-4)/1+1=29$ along each of the three spatial axes. Depth = number of filters = 16. So output = $29\times 29\times 29\times 16$.

---

🧾 Summary Table

---