Chapter 4 : Data Processing

Question 1

Where does data come from?

Answer 1

Agent’s sensors

Question 2

What does higher sampling frequency mean?

Answer 2

Less information missing

Question 3

What is aliasing?

Answer 3

When sampling, signals become aliases of each other and can’t reconstruct original signal anymore

Question 4

What is noise?

Answer 4

Obscures features in data

Question 5

What is artificat?

Answer 5

Makes it appear as a feature exists when it does not

Question 6

is DFT created on idea of in time any signal can be seen as a sum of sine functions

Answer 6

True

Question 7

What is PSD (Power Spectrum Density)?

Answer 7

Shows the power at each point of frequency. Co-related to squared amplitude of DFT.

Question 8

What is an modified Periodogram?

Answer 8

When the window size is not a vector of 1 x number of data points

Question 9

What is welch method?

Answer 9

1. Divide signal into segments
2. Take periodogram of each of the segments
3. Average all periodograms

Question 10

Purpose of welch method?

Answer 10

Smoother PSD with reduced variance

Question 11

What domain are filters applied?

Answer 11

Frequency domain

Question 12

What is frequency response?

Answer 12

Explains how a filter or system effects signals in frequency domain in terms of amplitude response and phase response

Question 13

What is FIR filter?

Answer 13

Impulse Response is finite because no feedback loop

Question 14

What type of filter is moving average filter?

Answer 14

Low pass filter

Question 15

Why normalize filter?

Answer 15

So the size of the output does not depend on size of input

Question 16

What is the FIR Output Filter Delay?

Answer 16

Output delay of symmetric FIR filters is

(Window size - 1) / 2

Question 17

Why do we need other filters other then moving average filter?

Answer 17

• Moving average filter needs to be large to remove a lot of noise
• A large filter causes more delay
• Transition band of moving average filter can be very large

Question 18

What is a matched filter?

Answer 18

FIR filter which tries to extract features from a known spatial signal

Question 19

What are problems with matched filter?

Answer 19

• Very sensitive to signal change

- Very fine-tuned for signals

Question 20

What type of output does median filter give?

Answer 20

Constant output (box version)

Question 21

What is the purpose of feature selection?

Answer 21

1. Achieves faster training
2. Select ML models with less complexities which leads to a less chance of overfit and easier to interpret
3. Better generalization can be achieved, accuracy can be improved

Question 22

What does variance-based feature selection do?

Answer 22

Trying to extract features with most information in them

Question 23

What does correlation-based feature selection do?

Answer 23

Tries to remove features that are very similar to other features

Question 24

What does univariate-based feature selection do?

Answer 24

Evaluate each feature individually with respect to output to see which ones are most important

Question 25

What does sequential-based feature selection do?

Answer 25

Uses greedy algorithm to pick local optimal features, uses both:

1. Forward-SFS algorithm
2. Backward-SFS algorithm

Question 26

What is the variable deletion techniques in missing data resolution?

Answer 26

Get rid of variables that have 50-60% missing data

Question 27

What is the curse of dimensionality?

Answer 27

As number of dimensions grow, the learning/search grows exponentially

Question 28

What can proper dimensionality reduction lead to?

Answer 28

1. Less complex processing
2. Faster processing
3. Easier visualization

Question 29

What are PCA algorithm steps

Answer 29

1. Normalize the data
2. Calculate the covariance matrix
3. Calculate eigenvectors and eigenvalues
4. Sort by highest-to-lowest by eigen values
5. Pick however many dimensions needed

Question 30

What happens if Pearson-correlation is 0?

Answer 30

The signals show no relationship to eachother

Question 31

What does Independent Component Analysis (ICA) do?

Answer 31

Used in source separation, maximizes the statistical independence

Question 32

What conditions are needed for ICA to perform?

Answer 32

1. Number of observes must be greater then or equal to number of sources
2. The different sources of information must be independent
3. Information must be additive

Question 33

Problems with ICA?

Answer 33

1. Order of outputs cannot be determined
2. Not perfect
3. Computationally expensive (iterative algorithm)

Question 34

Why is ICA called “blind” source separation (BSS)?

Answer 34

Due to the fact that there is not much particular information known about the sources