1. CIVL 7012/8012
Introduction to Machine Learning

2. What is Machine Learning?
Field of study that enables computers to learn by itself without being programmed.
Learning implies improving on certain task T, with respect to performance metric P based on training experience E
Some examples
Identifying spams in
T: Classify s as legitimate or spam
P: Percentage of accurate classifications
E: database
Detecting damages in structures
T: Detect the location and extent of damages in structure
P: Percentages of accurate detections
E: Visual and sensory data

3. Some more examples Autonomous vehicles
T: Vehicle driving on the road on its own using cameras and sensors
P: Distance travelled on its own until occurrence of human-judged error
E: Visual data and sensor collected by observing a human driver
Classifying soil using photographs
T: Classifying soil class and type
P: Percentages of accurate classifications
E: Visual data such as photographs
Identifying spoken words (How personal assistant such as Apple Siri, Amazon Alexa, etc. work)
T: Classify and identify words spoken by someone
P: Percentage of accurate identifications
E: Database of words spoken by people with different accents

4. Some more examples of tasks best solved by a learning algorithm
Classification and regression
Determining credit scores
Determining if someone is credit-worthy
Pattern recognition
- Such as shopping recommendations for items bought together. For example: mobile phones and protective cases)
Recognizing anomalies and outliers
- Such as unusual credit card transactions.
Forecasting and Predicting
- Stock prices in the future

5. How is ML different from traditional modeling?
Machine Learning
ML does not need human to program. The data is enough for the computer to recognize patterns and make predictions

6. ML components Every ML algorithms has three components:
Representation:
Evaluation
Optimization

7. Representation
The knowledge or the data should be presented in a formal language understood by the computer
Examples include decision trees, graphical models, neural networks, support vector machines, etc.
Depends on “features” or variables that represent the real work entity

8. Optimization
Learn a model from the features by minimizing certain criteria (e.g. the error)
Largely depends on ML algorithm used

9. Evaluation
Evaluate the program for different learners (ML algorithms) and distinguish good learners using some metric.
Similar to using different models to fit a data and choosing the best one in traditional modeling
Not all ML algorithms perform the same. Some might work very well at certain evaluations while performing poorly on others.

10. Types of learning Supervised learning
Training data included desired outputs (Remember regression?)
Unsupervised learning
Training does not include desired outputs
Used in finding hidden patterns in data (e.g. weather forecast)
Semi-supervised learning
Training data includes few desired outputs
Reinforcement learning
- “rewards” from “actions”

11. Supervised learning: Regression
Given (x1,y1), (x2,y2), …., (xn,yn) learn a function f(x) that can predict y given x.
Price of house as a function of area of house
Image source:

12. Supervised learning: Regression
Steps
Find the function f(x) best representing the data using training data which is called “Feature Engineering”.
Evaluate its performance using testing data
Look at other algorithms that can potentially work better
It is important to understand how the features interact or if the performance of ML algorithm is affected by feature type

13. Supervised learning: Regression
Some applications
Disease diagnosis based on symptoms of the patient
Face recognition based on picture of a person
Voice recognition based on the tone and pitch of an individual’s voice
Credit risk assessment based on credit history

14. Supervised learning: Classification
Given (x1,y1), (x2,y2), …., (xn,yn) learn a function f(x) that can predict y given x provided y is a categorical variable.
Example: Given past weather pattern, what will be the weather like tomorrow? Will it be sunny, or will it rain?
Classifying Class A and Class B which have similar features
Image source:

15. Unsupervised learning: Clustering
Given x1, x2, …., xn can predict hidden structure behind x’s
Clustering need not be spherical
Image source:

16. Reinforcement learning
Goal is to learn actions that maximize rewards
Lifelong learning without training and testing
Used in robotics, industrial manufacturing

17. Reinforcement learning
Example
Consider a self driving car stopped at a red-light signal. The light turn green and the car starts moving. Here the agent is the car and it is moving from one state to another. This actions produces some reward.
On the contrary if the car does not start and remains at the intersection despite green signal, it receives a traffic ticket. This is negative reward.

18. Reinforcement learning
Assume the world as a finite set of states
Each state has finite number of actions
Specific actions from a state results in a reward when executed and then we move to a different state

19. Evaluation metrics for ML classifiers
Accuracy alone is not a good measure of the accuracy of an algorithm.
Consider an algorithm detecting spams. Out of 10,000 s we know there are 100 spams.
Case 1: Any the algorithm sees is classified as Not-spam.
Accuracy for spam s only = 0/100
Case 2: Any the algorithm sees is classified as spam
Accuracy for spam = 100/100
Is the second algorithm better?

20. Precision and recall
Since accuracy alone does not provide a good idea of algorithm in many cases, we use two standard measures to evaluate quality of prediction.
Precision: Accuracy in terms of classifying relevant instances
Recall: Accuracy in classifying all relevant instances correctly
If nt is the number of true positives, nf is the number of false positives and ne is the number of false negatives
Precision: (nt)/(nt+nf)
Recall: (nt)/(nt+ne)

21. F1 score Used to evaluate the performance of classifiers
Combines precision and recall
Calculated as harmonic mean of precision and recall
F1=2*Precision*Recall/(Precision+Recall)

22. Example
Consider there are 100 known spam s in a dataset with 10,000 s. Find precision, recall assuming all the s the algorithm sees are classified as Not-spam.
nt=0, nf=0, nt=100
Precision=undefined
Recall=0

23. Errors There are two types of errors in ML algorithms:
Training error: Error made on training data
Testing error: Error made on testing data.
Unlike traditional modeling ML algorithms need to developed with generalization in mind in order to fit data unused in training.

24. Estimating errors
For traditional models we can estimate errors made the model in training within a confidence interval
ML algorithms however will strive to achieve perform very well on training data sometimes with zero error. Therefore traditional statistical method may not be applicable in evaluating ML algorithms.

25. Cross validation
Cross validation evaluates performance of supervised ML algorithms
Given a dataset with number of features or variables how do you decide which algorithms is the best?
We can
Train the algorithm to be more accurate (loses generalization)
Choose the most accurate algorithm
What about data that is not seen by the algorithm yet? How will it perform with such data?

26. Example
In this example, piece wise linear regression is a perfect fit for the dataset, but this maybe too perfect.
What about future data point though?
The first two are likely going to better fit for future point due to their better generalization.
Best algorithms are those which are accurate but do not lose generalization and can predict future uncertainties in data set better

27. Methods of Cross Validation
Train test:
Steps
Split the data set into training and testing data set. Usually the split is 75% training and 25% testing.
Develop a model using training dataset
Test model performance using training dataset.
Leave One Out Cross Validation (LOOCV)
Remove one observation from training data
Train ML algorithm on the remaining data
Test the algorithm with the removed data
Repeat it for all observations in the training data
LOOCV is cumbersome as all the observations in original dataset need to tested.
If there are 10 million observations then 10 millions training instances are needed.

28. Cross Validation K-fold cross validation
LOOCV is not suitable for huge datasets. K-fold validation is a more efficient approach
Steps
Split data into K partitions
Train K-1 partitions and test on remaining partition
Repeat this for each partition
Value of K depends on the data
10-fold cross validation is the most commonly used

29. Methods of Cross Validation
Method of validation used depends on the user.
Train test works best if the training data is not biased
LOOCV although validates all the data points is not scalable to large datasets
With LOOCV with exclusion of data can result in different data fit as the data changes constant