Lecture 4

Question 1

What is the difference between classification and regression tasks?

Answer 1

Classification is when the output is one of a finite set of values. Regression is when the output is a measured integer or real number.

• Classification = sunny/cloudy/rainy or true/false
• Regression = tomorrow’s temperature

Question 2

How do we choose a GOOD hypothesis space?

Answer 2

Choose the one that fits your model the best.

Question 3

What is induction?

Answer 3

Going from a specific set of observations to a general rule. We assume that we can apply our model to future cases (e.g., image recognition). NOTE: Inductive conclusions can be incorrect.

Question 4

What is a deductive conclusion?

Answer 4

Conclusions that are guaranteed to be correct if the premises are correct.

Question 5

What are the 3 types of Learning?

Answer 5

• Supervised Learning
• Unsupervised Learning
• Reinforcement Learning

Question 6

How do we choose a hypothesis space?

Answer 6

If you don’t have some prior knowledge about the process that generated the data, you perform exploratory data analysis to determine which hypothesis space is appropriate. (Or use trial and error.)

Question 7

How do we choose a good hypothesis from within the hypothesis space?

Answer 7

Look for a best-fit function for which each h(x_i) is close to y_i, which is the case if h generalizes well with the test set.

Question 8

How can we perform exploratory data analysis?

Answer 8

Examine the data with statistical tests and visualizations (such as histograms, scatter plots, box plots, etc.)

Question 9

What is the true measure of a hypothesis?

Answer 9

How well it handles input it has not yet seen (e.g., test set), not how it does on the training set.

Question 10

Define bias.

Answer 10

The tendency of a predictive hypothesis to deviate from the expected value when averaged over different training sets.

// or //

A model’s tendency to resist change. High bias == highly resistant to change (e.g. linear model).

Question 11

Define variance.

Answer 11

The model’s magnitude of change.

The amount of change in the hypothesis due to fluctuation in the training data.

Question 12

When is a hypothesis underfitting?

Answer 12

When it fails to find a pattern in the data.

Question 13

When is a hypothesis overfitting?

Answer 13

When it performs poorly on unseen data because it pays too much attention to a particular data set it is trained on

Question 14

Bias-variance tradeoff

Answer 14

A choice between:
1. more complex, low-bias hypotheses that fit the training data well
2. simpler, low-variance hypotheses that may generalize better

Question 15

What is Ockham’s Razor Principle?

Answer 15

Choose the simplest hypothesis that matches the data because there is often a bias-variance tradeoff.

Question 16

Regarding decision trees, what is the most important attribute?

Answer 16

The one that makes the most difference to the classification of an example.

Question 17

What is entropy?

Answer 17

A measure of the uncertainty of a random variable. (e.g., a loaded coin that always lands on heads has entropy of 0)

Question 18

What if we don’t have enough data to make all three of the data set splits?

Answer 18

You can use k-fold cross-validation.

Question 19

In your own words, explain how the Gradient Descent algo works.

Answer 19

There is a for-loop, and for every iteration you will modify your weights by slighty increasing or decreasing them, until you get to a point where it is optimal.

Question 20

Why would we want a machine to learn?

Answer 20

1. The designers cannot anticipate all possible future siutations
2. Sometimes the designers have no idea how to program a solution

Question 21

Classification learning problem

Answer 21

When the output is one of a finite set of values (e.g., boolean, sunny/rainy)

Question 22

Regression learning problem

Answer 22

When the output is a number (integer or real number)

Question 23

Supervised Learning

Answer 23

The agent observes input-output pairs and learns a function that maps from input to output (aka Labels). There is always an expected label for each processed input.

Question 24

Unsupervised Learning

Answer 24

The agent learns patterns in the input without any explicit feedback.

(The expectations are lower than in supervised learning)

Question 25

Reinforcement Learning

Answer 25

The agent learns from a series of reinforcements known as rewards and punishments. There are no explicit labels; instead, it has critics.

(e.g., chess game: agent won = reward, agent lost = punishment)

Question 26

What is the Ground Truth?

Answer 26

Output y_i - the true answer we’re asking our model to predict

Question 27

Why not let H be the class of all computer programs, or all Turing machines?

Answer 27

1. There is a tradeoff between the expressiveness of a hypothesis space and the computational complexity of finding a good hypothesis within that space
2. Simpler hypothesis spaces should be preferred because we want to use h after it’s learned

Question 28

Decision Boundary

Answer 28

A line (or a surface, in higher dimensions) that separates two classes.

(In linear regression, this is a straight line referred to as a linear separator)

Question 29

When is data linearly separable?

Answer 29

When it admits a linear decision boundary (linear separator)

Question 30

What is a parametric model?

Answer 30

A learning model that summarizes data with a set of parameters of fixed size.

Question 31

What is a non-parametric model?

Answer 31

A learning model that cannot be characterized by a bounded set of parameters.

(e.g., instance-based or memory-based learning)