AI Revision

Question 1

What is combinatorial explosion?

Answer 1

sIs where the possible combinations or permutations increases rapidly with the size of the input or problem

Question 2

What is the turing test? what is the objective? and provide one example of the test in real life

Answer 2

The turing test was a test designed by Alan Turing which consisted of an interrogator being connected to a person and a machine via a terminal and cannot see either of them, and its object would be to find out which candidate is the machine or human only by asking them questions. If the machine could fool the interrogator 30% of the time the machine is considered intelligent.

Real Life Examples: -> CAPTCHA | SIRI

Question 3

What is machine Learning?

Answer 3

The ability to learn without being explicitly programmed. Seen as part of AI, it builds a model based on sample data.

Question 4

What is supervised learning vs unsupervised learning

Answer 4

Is giving the machine learning model data which is supervised meaning that it relies on label input and output training data where as unsupervised learning processes unlabelled or raw data.

SIMPLE TERMS -> SUPERVISED -> Train a model on labelled data, goal is to predict the label of new instances. UNSUPERVISED -> Analysis of unlabelled data, goal is to uncover hidden patterns

Question 5

What are a few examples of Supervised and Unsupervised Learning: (2 each)

Answer 5

Supervised Learning ->
1. Handwriting Recognition
2. Traffic prediction

Unsupervised learning ->
1. Ad targeting ( group similar customer together and send adverts specific to those clusters)
2. Crime hotspot (Identify areas with high crime rates)

Question 6

What is clustering?

Answer 6

Is a technique used in unsupervised learning, it is given unlabelled dataset and similarity metric. And the goal is to find ‘natural’ partitioning, or groups of similar data points.

Question 7

What is association rules?

Answer 7

Is a technique used in unsupervised learning, it is given a set of transaction records containing items, and it’s goal is to produce dependency rules, to predict occurrence of one variable based on occurrences of another variable(s).

Question 8

What is data mining?

Answer 8

Analysis of large quantity of data to discover, valid, non-obvious data to the system, data that should be possible to act on the item and ultimately humans should be able to interpret the pattern/model in data.

Question 9

How is data trained in supervised learning?

Answer 9

Data is split into 2 types of data, (Training data) which is used to learn the parameters of the model and (Test data) which is used to test how well the model works

Question 10

What is pros and cons of decision trees

Answer 10

Pros -> Easy to interpret and explain decisions, reasonable training time.

Cons -> Small variations in data can lead to very different trees, only simple decision boundaries, over-complex trees - overfitting.

Question 11

How to construct decision trees

Answer 11

In decision trees there are a few things to consider,
1. The root node (largest information gain)
2. The Internal nodes so true / false are considered decision rules on features.
3. Leaf nodes predicted class label

Question 12

What is learning regression?

Answer 12

Supervised learning technique to estimate a relationship between an output variable Y (label) and input features(s) X | GOAL - Find line of best fit.

Question 13

Pros and cons of learning regression

Answer 13

Pros -> easy to interpret, fast to train, can work with limited training data.

Cons -> Only applicable if relationship is linear, relies on assumption, none of the features are highly correlated

Question 14

What is model evaluation?

Answer 14

How well does the model do on the validation set.

Question 15

What is model tuning?

Answer 15

Adjust model hyperparameters, e.g. -> maximum depth of decision trees

Question 16

What is overfitting?

Answer 16

Is where the model that performs well on the training data but poorly on unseen or test data because it has become too specialized to the training data and is not able to generalize well to new data.

Question 17

What is the concept of search space?

Answer 17

Set of all possible solutions to a problem

Question 18

Search trees construct

Answer 18

Root Node, Child/parent nodes / depth of a node / branches / leaves

States - nodes (possible states of the problem e.g. initial)
Search Space - all nodes in the tree (the set of all states reachable from the initial state)
Operators - Branches (a set of actions that move one state to another.)
Neighbourhood - All possible states reachable from a given state.
Goal Test - Test if the search reached a state that solves the problem
Path Cost - how much it costs to take a particular path

Question 19

What is the issues with search tree?

Answer 19

Combinatorial Explosion, Search efficiency (computationally expensive).

Question 20

What is the BFS Method

Answer 20

Expand root node first
Expand all nodes at level 1 before expanding level 2

Question 21

What is the DFS Method

Answer 21

Expand root node first
Explore one branch before exploring another branch

Question 22

What is the UCS Method

Answer 22

Expand root note first
Always remove the smallest cost node first in the queue

Question 23

What is learning rate?

Answer 23

Learning rate dictates how quickly the network converges by adjusting different scale to correct the weights.

Question 24

What is a linearly separable function?

Answer 24

Is a function which can be separated | Only linearly separable functions can be represented as a single layer NN, e.g. perceptron

Question 25

Explain the learning rate if too low and too high

Answer 25

Too low: Very small changes to weights → training will take a very long time however it is guaranteed to converge to correct weights if they exist
Too high: : Drastic adjustment to weights each time → may jump over a good combination of weights

Question 26

Neural network pros and cons

Answer 26

Pros - 1.Can learn more complicated decision boundaries.
2. Can handle a large number of features.
Cons - 1.Difficult to interpret, 2.difficult to design neural network architectures that work well for a problem - lots of hyper parameters to tune, 3.slow training time