Machine Learning

EE514 – CS535

Supervised Learning Overview,

Formulation, Train-test Split

Zubair Khalid

School of Science and Engineering

Lahore University of Management Sciences

https://www.zubairkhalid.org/ee514_2021.html
 Machine Learning: Overview
What is Machine Learning?
Given examples (training data), make a machine learn system
behavior or discover patterns

Data
Model

Algorithm f ( x)
Final output which
enables us to make
Given to us We need to design it predictions
 Machine Learning: Overview
Algorithms vs Model

- Linear regression algorithm produces a model, that is, a vector of values

of the coefficients of the model.

- Decision tree algorithm produces a model comprised of a tree of if-then

statements with specific values.

- Neural network along with backpropagation + gradient descent: produces

a model comprised of a trained (weights assigned) neural network.
 Machine Learning: Overview
Nature of ML Problems
1. Supervised Learning
The learning algorithm would receive a set of inputs along with the corresponding correct
outputs to train a model
Training Data

All labeled
Model Prediction
data
 Supervised Learning
Regression
Regression: Quantitative Prediction on a continuous scale
Examples: Prediction of
- Age of a person from his/her photo
- Price of 10 Marla, 5-bedroom house in 2050
- USD/PKR exchange rate after one week What do all these problems
- Efficacy of Pfizer Covid vaccine have in common?
- Average temperature/Rainfall during monsoon
Continuous outputs
- Cumulative score in ML course
- Probability of decrease in the electricity prices in Pakistan
- No. of steps per day

Predicting continuous outputs is called regression

 Supervised Learning
Classification
Classification: Given a data sample, predict its class (discrete)
Examples: Prediction of
- Gender of a person using his/her photo or hand-writing style
What do all these problems
- Spam filtering
have in common?
- Object or face detection in a photo
Discrete outputs: Categorical
- We will be back on Campus on Feb 1
- Temperature/Rainfall normal or abnormal during monsoon Yes/No (Binary Classification)
- Letter grade in ML course
Multi-class classification:
- Decrease expected in electricity prices in Pakistan next year multiple classes
- More than 10000 Steps taken today

Predicting a categorical output is called classification

 Supervised Learning Setup
Nomenclature

In these regression or classification problems, we have

- Inputs – referred to as Features
- Output – referred to as Label
- Training data – (input, output) for which the output is known and is
used for training a model by ML algorithm
- A Loss, an objective or a cost function – determines how well a trined
model approximates the training data
- Test data – (input, output) for which the output is known and is used
for the evaluation of the performance of the trained model
 Supervised Learning Setup
Nomenclature - Example
Predict Stock Index Price

- Features (Input)
- Labels (Output)
- Training data

?
?
?
 Supervised Learning Setup
Formulation

?
?
?
 Supervised Learning Setup
Formulation

Regression:

Classification:
 Supervised Learning Setup
Example

Data of 200 Patients:

- Age of the patient
- Cholesterol levels
Model (h) Prediction of Oxygen Saturation
- Glucose levels
- BMI
- Height
- Heart Rate
- Calories intake
- No. of steps taken
 Supervised Learning Setup
Example

Model (h) Prediction

MNIST Data:
- Each sample 28x28 pixel image
- 60,000 training data
- 10,000 testing data
 Supervised Learning Setup
Learning
 Supervised Learning Setup
Hypothesis Class

Q: How?
A:
 Supervised Learning Setup
Q: How do we evaluate the performance?
A:

Loss Function
 Supervised Learning Setup

0/1 Loss Function:

Zero-one loss is defined as

Interpretation:
- Note normalization by the number of samples. This makes it the loss per sample.
- Loss function counts the number of mistakes made by hypothesis function D.
- Not used frequently due to non-differentiability and non-continuity.
 Supervised Learning Setup

Squared Loss Function:

Squared loss is defined as (also referred to as mean-square error, MSE )

Interpretation:
- Again note normalization by the number of samples.
- Loss is always nonnegative.
- Loss grows quadratically with the absolute error amount in each sample.
Root Mean Squared Error (RMSE):
RMSE is just square root of squared loss function:
 Supervised Learning Setup

Absolute Loss Function:

Absolute loss is defined as

Interpretation:
- Loss grows linearly with the absolute of the error in each prediction.
- Used in regression and suited for noisy data.

* All of the losses are non-negative

 Supervised Learning Setup
Learning

(Optimization problem)

Recall

Q: How can we ensure that hypothesis h will give low loss on the input not in D?
 Supervised Learning Setup

Interpretation:
- 0% loss error on the training data (Model is fit to every data point in D).
- Large error for some input not in D
- First glimpse of overfitting.
Revisit:
Q: How can we ensure that hypothesis h will give low loss on the input not in D?

A: Train/Test Split
 Supervised Learning Setup
Generalization: The Train-Test Split

How to carry out splitting?

You can only use the test dataset once after deciding on the model using training dataset
 Supervised Learning Setup
Learning (Revisit after train-test split)

Evaluation
 Supervised Learning Setup
Generalization loss
 Supervised Learning Setup
Generalization: The Train-Test Split

Q: Idea:
Validation data is used evaluate the loss for a function h that is
determined using the learning on the training data-set. If the loss
on validation data is high for a given h, the hypothesis or model
needs to be changed.
 Supervised Learning Setup
Generalization: The Train-Test Split

More explanation* to better understand the difference between validation and test data:

- Training set: A set of examples used for learning, that is to fit the
parameters of the hypothesis (model).

- Validation set: A set of examples used to tune the hyper-parameters of

the hypothesis function, for example to choose the number of hidden
units in a neural network OR the order of polynomial approximating
the data.

- Test set: A set of examples used only to assess the performance of a

fully-specified model or hypothesis.

Adapted from *Brian Ripley, Pattern Recognition and Neural Networks, 1996
 Supervised Learning Setup
Generalization: The Train-Test Split (Example)

Cross validation simulates multiple train-test splits on the training data

 Supervised Learning Setup
Reference:

• https://www.cs.cornell.edu/courses/cs4780/2018fa/

• CB: sec 1.1

• HTF section 2.1

• KM: sec. 1.1, 1.2

Feedback: Questions or Comments?

Email: [email protected]