Machine learning

Question 1

What is machine learning?

Answer 1

Computer system or software learns by itself by developing models and training them to predict future outputs

Question 2

What is supervised learning?

Answer 2

Infers a function from labelled training data

Question 3

What is unsupervised learning?

Answer 3

Infers a function from unlabelled data

Question 4

What is reinforcement learning?

Answer 4

Learns over time via trial and error using feedback
- Award from actions

Question 5

Give some examples of supervised learning

Answer 5

• Linear regression
• Decision tree
• Artificial neural networks

Question 6

Give some examples of unsupervised learning

Answer 6

• Clustering
• Association rules

Question 7

What is top-down machine learning?

Answer 7

Model different functions and wire them together
- Deduction

Question 8

What is bottom up machine learning

Answer 8

Give the system lots of data so it can discover the concepts by itself
- Induction

Question 9

How does supervised learning work?

Answer 9

1. Data pre-processing
2. Partition data into training and testing
3. Train model

Question 10

How does unsupervised learning work?

Answer 10

1. Data pre-processing
2. Clustering or association technique

Question 11

How does clustering work?

Answer 11

1. Choose number of clusters K
2. Initialise K cluster centroids randomly
   Repeat steps 3 and 4
3. Assign each data point to the nearest cluster
4. Update centroids by computing the mean of all the data points assigned
5. Output final cluster assignments and centroids

Question 12

How does association work?

Answer 12

1. Discover correlation between two or more variables
2. Produce dependancy rules to predict occurrence of x with y

Question 13

What are the three pillars that machine learning is built from

Answer 13

1. Models and algorithms
2. Powerful and cheap computation
3. Massive data warehouses

Question 14

What is data mining?

Answer 14

The exploration and analysis of large quantities of data to discover valid, novel, useful and understandable patterns in data

Question 15

What is the difference between machine learning and data mining?

Answer 15

• Machine learning predicts using models
• Data mining explains patterns

Question 16

What is regression?

Answer 16

A relationship between variable Y and variable X

Question 17

How do we describe a linear regression model?

Answer 17

An underfitted model
A good model
An overfitted model

Question 18

What is meant by underfitting a model?

Answer 18

A model which doesn’t caputre any logic
- High loss
- Low accuracy

Question 19

What is meant by a good model?

Answer 19

Caputres the underlying logic of the dataset
- low loss
- high accuracy

Question 20

What is meant by an overfitted model?

Answer 20

Caputures all the noise, so “misses the point”. Over complex with lots of parameters
- low loss
- low accuracy

Question 21

How may overfitting occur?

Answer 21

Training data size is too small
-> take more samples (could use deeplearning GANs to do this)

Question 22

How may underfitting occur?

Answer 22

Model is too simple, too little parameters
-> more training time or input features

Question 23

What are the advantages of regression?

Answer 23

• Short training time
• Easy to interpret
• Easy to implement

Question 24

What are the disadvantages of regression?

Answer 24

• Sensitive to noise and outliers (overfitting)
• Cannot handle complicated relationships (linear only)

Question 25

What two data type can a label be?

Answer 25

• Categorical label
• Continuous label

Question 26

What is categorical data

Answer 26

Data which can be sorted into groups/ categories
- Classification
- good or bad

Question 27

What is continuous data?

Answer 27

Data which can take any value
- Regression
- probability

Question 28

What are the components of a decision tree?

Answer 28

• Internal nodes
  Features (decision variables, inputs)
• Branches
  Course of decision or action
• Leaf nodes
  A predicted class label (output)

Question 29

How do you train a decision tree?

Answer 29

Iteratively partition the decision space
- What values to split on?
- What features to split on?

Question 30

What are the advantages of decision trees?

Answer 30

• Reasonable training time
• Easy to interpret
• Easy to implement
• Can handle large number of features

Question 31

What are the disadvantages of decision trees?

Answer 31

• Over-complex trees lead to over fittting
• Cannot handle complicated relationships
• Only simple decision boundaries
• Problems occur when there is lots of missing data

Question 32

What is a neural network?

Answer 32

A set of neurons connected by directed, weighted edges

Question 33

What are the advantages of a neural network?

Answer 33

• Can handle a large number of features
• Can be more accurate
• Can learn more complicated class boundaries

Question 34

What are the disadvantages of a neural network?

Answer 34

• Overfitting of data
• Hard to implement
• Slow training time
• Hard to interpret

Question 35

Give a practical example of overfitting

Answer 35

Develop a model to recognise letters from peoples handwriting. Model is trained with a small sample of people so cannot recognise other peoples handwriting

Question 36

Give a practical example of underfitting

Answer 36

Develop a model to predict housing costs with number of bedrooms. There are many other factors like number of bathrooms, location etc so would not predict correctly.