03 - Datasets

Question 1

Famous datasets

Answer 1

• MNIST Digits
• ImageNET (WordNet hierarchy)
• BRATS (brain tumors)

Question 2

Types of datasets

Answer 2

• classification: 1 img = 1 value
• regression: 1 img = predict 1 or more decimal values
• segmentation: 1 pixel = 1 value
• detection: group of pixels = 1 obj

Question 3

Common issues with datasets

Answer 3

• imbalance
• too homogeneous
• bias
• too few examples

Question 4

Augmenting datasets

Answer 4

• transform = shift, zoom, rotate, intensity, color, shear
• other = noise, blurring
• /!\ how much distortion is acceptable

Question 5

Baseline

Answer 5

= simple, easily implemented and understood model that illustrates the pb and the ‘worst case scenario’ for a model that learns nothing
- nearest neighbor (1 or k) is usually a good baseline