06 Representation Learning

Question 1

What is representation learning

Answer 1

classification, detection, segmentation
eg. detect what is in the picture, extract readable summary from articles, sentimental analysis

1. discovering the nature of data and learning the representation
2. mapping high dimensional data into low dimensional data
3. make subsequent machine learning task easier

Question 2

what are performance measurements

Answer 2

1. classification accuracy
   (true positive + true negative) / true positive + true negative + false positive + false negative
2. recall
   true positive/ (true positive + false negative)
3. precision
   true positive/ (true positive + false positive)
4. harmonic F-score
   2 * (precision * recall) / (precision + recall)

Question 3

what is projection

Answer 3

project data (D dimensions) into lower number of (M dimensions)
x = N * D
W = M * D (eg. where to cast the projection)
z = x * w

Question 4

what is principle component analysis

Answer 4

a method for choosing W
- finds the columns of W one at a time by defining M

Question 5

what is an autoencoder

Answer 5

neural networks trained to reconstruct output based on inputs

encoder -> encode and compress information through network
code/ bottleneck -> compressed representation of input data
decoder -> reconstruct output

Question 6

what is ADAM

Answer 6

• adaptive moment estimation

Question 7

PCA vs auto encoders

Answer 7

pca
- straight line, linear representation of data

autoencoder
- non linear dimensionality reduction
- high efficiency
multiple transformation