1 kNN Classification

Question 1

What is DL?

Answer 1

DL is ML with the addition of the learning process being handled by a NN.
- Layered architecture of computational units.
- High-capacity models possible, called ANN’s or DNN’s
- Restricted to DNN (mores than 1 hidden layer)

Question 2

What are S+C cells?

Answer 2

Simple → Complex → Hypercomplex cells

Question 3

What is Marr’s theory on visual cortex?

Answer 3

• s+c cells compute a primal skecth
• The brain computes a 2.5D sketch using texture
• A 3D model is computed

Question 4

What is texture based descriptors?

Answer 4

Methods that cann identify and use these patterns for tasks like image recognition or classification.
- SIFT, SURF, HoG

Texture - visual patterns

Question 5

What is Representation learning?

Answer 5

Learn both the representation an dhow to produce the required output.

Refers to ML techniques focused on learning and extracting meaningful representations or features from raw data.

Question 6

What are well known datasets?

Answer 6

• MNIST → small and fast
• PASCAL → classification, detection and more
• ImageNet → classification + localization, deection
• COCO → detection, segmentation

Both YOLO and SSD is able to do real time detection.

Question 7

What types of ML tasks are there?

Answer 7

• Classification
• Regression
• Compression
• Clustering
  …

Question 8

What advances does DL have?

Måske overflødig?

Answer 8

• Sequence (e.g. speech) processing
• Machine translation
• Image segmentation
• Image captioning

Question 9

What does Bayes rule say?

Answer 9

p(C_k│x)=( p(x│C_k )p(C_k) ) / ( p(x) )

Question 10

What is AI?

Answer 10

• The concept of “human-like” machines.
• Many, many subdisciplines

Question 11

What is ML?

Answer 11

• Focus on data
• Instead of being designed, the machine learns a model from the data.

Question 12

What do texture based (local) descriptors do?

Answer 12

Go from image gradients to keypoint descriptor

Question 13

How does testing work in the kNN classifier?

Answer 13

For each data point (image), find the k nearest training examples.
Decide on label by majority vote.

Question 14

How does training work in the kNN classifier?

Answer 14

Simply store the images and labels.

Question 15

What parameters are needed for kNN?

Answer 15

Data point x → given class C → result
K nearest datapoints → hyperparameter
Total number of training data points N → Countable

Question 16

For data point x, how do you decide on its class C?

Answer 16

For each k’th class calculate:
p(C_k|x)
The winning class is the one with the highest probability

Question 17

How do you calculate the probability of x conditioned on class k?

Answer 17

p(x|C_k)= K_k / N_k

K_k → Number of hits in that class, N_k → Total number rof entities in that class

Question 18

How do you calculate the unconditioned probability of x?

Answer 18

p(x)= K / N

K → hyper parameter, N → Total number

Question 19

What is class prior?

Answer 19

Unconditional probability of C_k

Probability of encountering instances of class C_k in the absence of specific information about the features.

Question 20

How do you calculate the unconditional probability of C_k?

Answer 20

p(C_k)= N_k / N

K_k → Total number of entities in that class
N → Total numberr of entities

Question 21

What is the final solution of kNN?

Answer 21

p(C_k│x)=( p(x│C_k )p(C_k) ) / ( p(x) )=K_k/K

Question 22

What are some rules you should know?

Answer 22

• Dont make k too large (biggest groups wins always)
• k should be uneven number
• Flip a coin if tied
• Low values of 1 or 2 can be noisy (outliers)
• Use Elbow method to compare the accuracies for different values of K.
• Use euclidean distance

Question 23

Example of image tasks?

Answer 23

• Image classification
• Single-object localization
• Object detection

Question 24

WHAT IS THE CANNY EDGE detector?

Answer 24

• Very early computational approach to vision.
• Find string intensity discontinuties in grayscale

Question 25

What does the Harris corner detector do?

Answer 25

Uswe local pixel statistics to find cornsers instead of edges.

Question 26

What does KNN do?

Answer 26

Make decision boundaries.