Supervised And Un-Supervised Learning

Question 1

Supervised learning

Answer 1

Task: learns from labelled training data (each sample/instance/data point consists of an input and a desired output)
Goal: find/train a model to match the output

Algorithms: linear regression, decision tree, ANN

Question 2

Unsupervised learning

Answer 2

Task: learns form unlabelled data to describe hidden structure/pattern
Goal: task dependent (eg. Group data points, find the distribution)

Algorithms: K-means, Apriori algorithm

Question 3

Reinforcement learning

Answer 3

Learns over time via trial and error using feedback/awards from actions
Goal: learn a policy to make decisions (to maximise rewards)

Algorithms: deep neural networks, Q-learning, deep Q-networks

Question 4

Training data is used to…

Answer 4

Learn the parameters of the model

Question 5

What is the process for supervised learning

Answer 5

Prepare the training data
There should be input and output datasets
Partition the dataset into two subsets
70% training - 30% testing

Train the model on the relevant training data

Test and evaluate the models on the test set
How accurate the trained model works on the test set

Question 6

What is the test set used for

Answer 6

How to generalise to new data
Unbiased estimate of how well the model works

Question 7

What does data pre-processing involve

Answer 7

Processing data into a suitable format for machine learning

Use pandas data frame library
Improve model accuracy
Reduce noise and bias

Data collection and data transformation

Question 8

What is data collection

Answer 8

Gathering relevant data from sources

Question 9

What is data transformation

Answer 9

Encode categorical values
Normalise numerical data
Split data (training, testing)

Question 10

What is data cleaning

Answer 10

Handling missing data, imputation
Removing, outliers and duplicates

Question 11

What is clustering

Answer 11

Unsupervised learning

Group similar objects
Given: un-labelled dataset D and similarity/distance metric
Goal: find natural partitioning, or groups of similar data points

Question 12

K-means clustering

Answer 12

Choose number of clusters ‘k’

Initialise k cluster centroids randomly
Repeat until a max number of iterations
A. Assign each data point to the nearest centroid (based on distance metric)
B. Update the centroids by computing the mean of all their data points assigned

Output the final cluster assignments and cnetroids

Question 13

Association Rules

Answer 13

Unsupervised learning

Discover correlation between two or more given variables
Given: a set of records containing items
Goal: produce dependency rules, to predict the occurrence of variable X with variable(s) Y
Algorithms: Apriori algorithm, frequent pattern growth

Question 14

Steps of building a k-means clustering model (unsupervised learning example)

Answer 14

Prepare the training
Data preparation
Understand the problem (which approaches will you use)
Read in the dataset
Observe/visualise the dataset

Training a k-means model
- Prepare the training data x
- Create a k-means clustering model

Results
What are the k centres
Show the centroids

Question 15

Categorical vs continuous supervised learning

Answer 15

These are labels
If the data is categorical we use classification
If the data is continuous we use regression

Question 16

Regression models

Answer 16

Show the relationship between y and x

Question 17

Regression training

Answer 17

Based on the given data find the function that minimises its mean squared error to ‘fit’ the samples

Question 18

Polynomial linear regression cons

Answer 18

Can run into issues of overfitting, fits too well with noise/outliers/errors

Question 19

Pros and cons of regression

Answer 19

Short training time
Easy to interpret
Easy to implement

Sensitive to noises and outliers
Cannot handle complicated relationships

Question 20

What is classification

Answer 20

Learn to predict to which set an instance belongs based on pre-labels (classified) instances

Question 21

Decision tree - internal nodes

Answer 21

Decision rules on features (decision variables, input)

Question 22

Decision tree - branches

Answer 22

Course of decision or action

Question 23

Decision tree - leaf nodes

Answer 23

A predicted class label (output)

Question 24

Pros and cons of decision tree

Answer 24

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

Only simple decision boundaries
Problems with lots of missing data
Cannot handle complicated relationships
Over-complex tree : overfitting

Question 25

What does ANN stand for

Answer 25

Artificial neural networks

Question 26

Neural Networks history

Answer 26

1943 - first neural network Combine simple inputs 1949 - first learning rule Added fixed weights 1950s-60s - Perceptron Converge to correct weights learning —> thinking The AI winter Mid 80s solved non-linear separable problem with multi layered networks

Question 27

Basic model of a neural network

Answer 27

Neurons connected by direct weighted paths Each neuron has a threshold t Weighted sum input to the neuron has to be greater than or equal to the threshold for the neuron to fire

Question 28

Training neural networks

Answer 28

Epoch - the entire training set feed to the neural network The AND function - an epoch consists or four sets of inputs (patterns) to feed into the network ([0,0],[0,1],[1,0],[1,1]) Training value Y - the value that we require the network to produce (Supervised learning) Error, err - the output by the network O differs from the training value T LR = learning rate Xi = inputs to the neuron Wi = weight from input to output While epoch produces an error Check nest input/patterns from epoch Err = Y - O If Err <> 0 then Wi = wi + LR * Xi * Err End if End while

Question 29

Pros and cons of ANN

Answer 29

Can learn more complicated class boundaries Can be more accurate Can handle large number of features Hard to implement - trial and error for choosing parameters and network structure Slow training time Can overfit the data - find pattens in random noise Hard to interpret

Question 30

What is deep learning

Answer 30

Based on artificial neural networks, multiple layers of processing are used to extract progressively higher level features from data

Question 31

Excitatory vs Inhibitory weights

Answer 31

Excitatory weights encourage the neuron to fire Inhibitory weights prevent the neuron from firing

Question 32

What is linearly separable data

Answer 32

Where data points (classes) can be separated using a linear boundary Only linearly separable functions can be represented by a single layer NN