Introduction

Question 1

What is image classification?

Answer 1

Determine what object is in the image, ouput e.g. “Bird”.

Question 2

What is object detection?

Answer 2

Detect where in the image a spesific object is, ouput e.g. center pixel.

Question 3

What is object localization?

Answer 3

Detect all “objects” in a image, can be of different types. Output e.g. bounding box.

Question 4

What is Object recognition?

Answer 4

Recognize what type of objects are in the image, differs from image classification in that we might have several objects in one image.

Question 5

What is semantic segmentation?

Answer 5

Segment objects pixelswise and add a label to each object.

Question 6

What is image captioning?

Answer 6

Describes images with text, or finds images in a database using text as input.

Question 7

What is image registration?

Answer 7

Find out how overlapping images are related.

Question 8

What type of cost function do we usually use for logistic regression?

Answer 8

cost =

• log(h(theta)) iff y = 1
• log(1-h(theta)) iff y = 0

Question 9

What is lasso and ridge penalty?

Answer 9

Lasso = L1 norm = lambda * sum abs(theta_i)
Ridge = L2 norm = lambda * sum (theta)^2

Question 10

What is the elastic net penalty?

Answer 10

A combination of L1 and L2 penalty.

Question 11

How can we fix a high variance in the loss (overfitting)?

Answer 11

1. Add more training data
2. Increase regularisation
3. Use less complex models
4. Use fewer features

Question 12

How can we fix a high bias in our trained models (underfitting)?

Answer 12

1. Decrease regularisation
2. Use more complex models
3. Use more features