Common network architectures

Question 1

Inception: State the multiplication cost for one convolution path and why bottlenecks help.

Answer 1

Mults = H·W·C·K²·F. 1×1 bottlenecks cut C, slashing FLOPs before larger kernels.

Question 2

ResNet: Write the residual output equation and its purpose.

Answer 2

y = F(x) + x. Lets the block learn a residual; identity mapping is easy so very deep nets avoid degradation.

Question 3

Precision formula and plain‑language meaning.

Answer 3

Precision = TP / (TP + FP): fraction of predicted boxes that are correct.

Question 4

Recall formula and plain‑language meaning.

Answer 4

Recall = TP / (TP + FN): proportion of ground‑truth objects the detector finds.

Question 5

Difference between precision and recall in one sentence.

Answer 5

Precision asks ‘How many predictions are right?’; recall asks ‘How many real objects did I catch?’

Question 6

Why use 1×1 convs before 3×3/5×5 in Inception modules?

Answer 6

They reduce channel depth, lowering compute while keeping representational power.

Question 7

How do residual skips help gradient flow?

Answer 7

They create a direct path so gradients bypass deep stacks, making optimisation easier for very deep networks.

Question 8

Give the IoU formula and common TP threshold.

Answer 8

IoU = overlap area / union area; a detection is TP if IoU ≥ 0.5.

Question 9

Define AP and mAP briefly.

Answer 9

AP = area under precision‑recall curve for one class; mAP = mean AP over all classes.

Question 10

Conv output size formula (no padding) and example for 28×28, F=3, S=2.

Answer 10

H_out = ⌊(H_in−F)/S⌋+1 ⇒ ⌊(28−3)/2⌋+1 = 13.