CSSC ML Workshop

Intro to
Machine Learning!
(Starting 5:10)

Agenda
● Speaker Intros
● How does a machine learn?
● Buzzword clarifications
● Supervised vs Unsupervised vs Reinforcement Learning
● ML Models
● Neural Networks (Structure, Activation and Cost Functions, Gradient Descent,
Backpropagation)
● Code Exercise Setup
● Code Exercise Walkthrough
● Q&A

Let’s Start with a Question
What does it mean for a
Machine to LEARN?

How does a Computer LEARN?
● First we have to gather relevant data
● Then feed the data into our model, sample by sample
● We then compare our models output to the output we want and we adjust the
many parameters (weights, biases) within the model so that it can output a
more accurate answer on future iterations
● After repeated iterations, the model’s answers with each inputted sample will
get better and better
● This is called Training the model
● This is how a computer learns (with supervised learning)

Buzzwords
Can you think of any you hear often?

AI vs ML vs Deep Learning vs Data Science
● AI
○ Scientific field aimed at making machines act like
humans
● ML
○ Subset of AI focused on using computational
models to allow computers to perform intelligent
tasks without being explicitly programed to do so
● Deep Learning
○ Subfield of ML dedicated to the engineering,
research, and use of neural networks
● Data Science
○ Field of statistics aimed at trying to find useful
patterns and structures within data

Supervised Learning
● Uses LABELED training data
● Each input in the data set has a corresponding output allowing us to measure
how well our model is performing
● Usually used to solve two main types of problems:
○ Classification problems
○ Regression problems

Supervised Learning Cont.
Classification: Regression:

Unsupervised Learning
● Uses non-labeled data
● There are no output labels to allow us to check how our model is doing
● Usually used to find patterns, structures, relationships within your data
● Used to solve problems like:
○ Clustering
○ Anomaly Detection
○ Association Rule Mining
○ etc.

Reinforcement Learning
● Involves an agent that can interact
with the world in ways that can give
a reward
● The agent is tasked with exploring
its environment to find ways to
maximize the reward
● Does not require a data set
● Similar to how we would train a dog
● Can be used to for:
○ Robotics
○ Game playing (Chess, Go, Dota 2, etc)
○ Self driving

ML Models
● A machine learning model is an algorithm (or set of algorithms) and
parameters that the computer uses to learn
● The model takes in input data, gives an output, and adjusts its parameters to
make itself more accurate
● The models we use are problem specific i.e. some models are good at certain
problems while others are not
● Model examples:
○ Linear Regression: Fitting a linear equation to data, used in predicting stock prices, home
prices, etc
○ K-Nearest Neighbors (KNN): Clusters data points into groups based on similar features, used
for classification and regression
○ Decision Trees: Tree-like structure, used for both classification and regression
○ Neural Networks: Used in deep learning models for image recognition, NLP, etc.

Neural Networks: Structure
● Structure inspired by the human brain
hence the term ‘neural’
● Each circle is called a ‘neuron’
● Neurons are arranged in columns
called layers
● Each neuron in one layer is connected
to every neuron in the following layer
● Usually always 1 input and 1 output
layer
● We can have many hidden layers
(depth)
● There’s typically many more neurons
in in the layers than shown in the
diagram

Neural Networks: Structure Cont.
● Neurons can ‘fire’ with values from
0-1
● Specific groups of neurons firing at
the same time cause neurons in
the next layer to fire
● Each connection between neurons
has a ‘weight’ describing how much
a neuron activation affects the
activation of the adjacent neuron
● Each neuron also has a value to
act as a ‘bias’ against activation so
a neuron only fires after a certain
threshold

Neural Networks: Activation Functions
● The activation function is a function that
gives the activation for a single neuron
● It takes in as input, the sum of the
weighted activations of all the neurons in
the previous layer + bias
● If Z is the input to the activation function,
then we get:
Z = w1x1 + w2x2 + w3x3 + ... + b
Where wn and xn is the weight and
activation for input neuron n respectively,
and b is the bias for the current neuron
Example: Rectified Linear Unit
activation function (ReLU)

Neural Networks: Activation Functions Cont.
● Each Neuron can be thought of as
a function
● Each neuron takes in the weights
and activations of the neurons in
the previous layer then adds in its
bias and outputs its own activation
value using the activation function
● Multiple different activation
functions can be used for different
problems

Neural Networks: Cost Functions
● The ‘Cost function’ is used to quantify how well our models outputs match the
correct output values
● Can also be referred to as ‘Loss Functions’
● We use the Cost function to see how well our model is doing and use that
metric to guide the optimization of the parameters (weights, biases) of our
model
● Examples of Cost Functions:
○ Mean Squared Error (MSE)
○ Binary Cross-Entropy Loss (Log Loss)
○ Negative Log-Likelihood (NLL)

Neural Networks: Cost Functions Cont.
● Cost functions take as input ALL the weights and biases of the neural network
● Then run all the training data on the model using the given weights/biases
● Then output a single value called the ‘Cost’ which represents how good/bad
our model is at outputting the correct answers
● A higher cost value implies a larger difference between the models outputs
and the correct outputs
● A lower cost value implies a smaller difference between the models outputs
and the correct outputs
● The goal is to minimize the Cost function by adjusting the weights and biases
of our model
● Minimizing Cost function == LEARNING

Neural Networks: Minimizing Cost
● If we want our model to perform
better, we need to minimize our
Cost function
● Similar to high school Calculus
‘finding minimum of the function’
type of questions
● Unfortunately when we’re dealing
with a large number of inputs to our
function, it’s much harder to find
local minima
● We require another approach

Neural Networks: Minimizing Cost Cont.
● One common approach is called
‘Gradient Descent’
● Gradient Descent involves
finding a direction that we can
move to that lowers cost
● Once we find the Gradient, we
use it to go backwards in our
neural network to update the
weights and biases
● Going backwards in our network
to update the weights and biases
is called ‘Backpropagation’
● We repeat this until we get to the
local minimum

PyTorch
● What is PyTorch?
● An open-source machine learning framework.
● Developed by Facebook's AI Research lab (FAIR).
● Key Features:
● Dynamic computational graph.
● Pythonic and ﬂexible.
● Excellent support for deep learning.
● Popularity:
● Widely adopted in research and industry.
● Preferred by researchers and practitioners for its ﬂexibility.
● Use Cases:
● Deep learning research.
● Computer vision, natural language processing, reinforcement learning, and more.

Handwritten Digit Recognition
● We will be using the MNIST data set for testing and training
○ The MNIST data set contains lots of handwriting samples
○ Each image is 28x28 in dimension *
● This is supervised learning for a classification task.
● There are six tasks for you to complete, 4 of them in main.py and 2 in extract.py
● Please work with others but do not give the answers to everyone!!
● Feel free to ask for help.

Code Exercise Setup
1. Open Powershell (Windows) or Terminal (Mac) AS ADMIN if possible
2. Enter ‘python --version’ and make sure version >=3.9
3. If you have an older version, download and install latest python version
4. In your Powershell/Terminal, Enter ‘pip install poetry’ and make sure it downloads
successfully
5. Go to the link https://t.ly/PoBkE -> Click ‘Code’ -> Click ‘Download ZIP’ and unzip the file
6. Copy the file path that goes to files within the unzipped digit-recognition-main folder
Windows example file path:
C:UsersNameDownloadsdigit-recognition-maindigit-recognition-main
7. In your Powershell/Terminal, Enter ‘cd FILE_PATH’ where FILE_PATH is the path you
copied earlier
8. Enter ‘poetry install’ and make sure it completes successfully
9. Enter ‘poetry shell’
10. Then enter ‘python main.py --help’ to get the full list of commands

CSSC ML Workshop

More Related Content

Similar to CSSC ML Workshop

More from GDSC UofT Mississauga

Recently uploaded

CSSC ML Workshop