Introduction to Machine Learning

Maggie Makar

Presidential Postdoctoral Fellow/Assistant Professor

Electrical Engineering and Computer Science

My work: clinical decision support through…

Prediction

Does this X-ray show signs of Pneumonia?

Causal inference

Will chemotherapy extend the patient’s life expectancy?

Start treatment

Chemo

No action

Palliative care

Does this X-ray show signs of Pneumonia?

Will chemotherapy extend the patient’s life expectancy?

Yes

No

Yes

No

What is Machine learning?

The study of patterns to create rules that generalize

What is Machine learning?

The study of patterns to create rules that generalize

What is not Machine learning?

Artificial Intelligence

Machine learning

What is Machine learning?

The study of patterns to create rules that generalize

What is not Machine learning?

Artificial Intelligence

Machine learning

“Prescribed” rules

E.g., Boolean and Fuzzy logic

If rice is brown and heat is 100 🡪 stop

​

​

If rice is white and heat is 100 🡪 lower heat

What is Machine learning?

The study of patterns to create rules that generalize

What is not Machine learning?

Artificial Intelligence

Machine learning

“Prescribed” rules

Learned rules

E.g., Boolean and Fuzzy logic

Don’t know what the rule is, but we try to ”learn” it from data

If rice is brown and heat is 100 🡪 stop

​

​

If rice is white and heat is 100 🡪 lower heat

Types of Machine Learning

• Supervised learning:
• Input is labeled data, goal is to generate a prediction for new examples
• E.g., Using patient characteristics to predict future adverse outcomes, using chest X-rays to detect presence of Pneumonia

​

​

​

Types of Machine Learning

• Supervised learning:
• Input is labeled data, goal is to generate a prediction for new examples
• E.g., Using patient characteristics to predict future adverse outcomes, using chest X-rays to detect presence of Pneumonia
• Unsupervised learning:
• Input is unlabeled data, goal is to discover new structures/clustering
• E.g., subtyping autism trajectories

​

​

Types of Machine Learning

• Supervised learning:
• Input is labeled data, goal is to generate a prediction for new examples
• E.g., Using patient characteristics to predict future adverse outcomes, using chest X-rays to detect presence of Pneumonia
• Unsupervised learning:
• Input is unlabeled data, goal is to discover new structures/clustering
• E.g., subtyping autism trajectories
• Reinforcement learning:
• Input varies (online vs offline RL). In offline RL state, action pairs and a reward signal. Goal is to identifying sequence of actions to optimize long term rewards
• E.g., Optimal strategy for sepsis treatment

​

​

Types of Machine Learning

• Supervised learning:
• Input is labeled data, goal is to generate a prediction for new examples
• E.g., Using patient characteristics to predict future adverse outcomes, using chest X-rays to detect presence of Pneumonia
• Unsupervised learning:
• Input is unlabeled data, goal is to discover new structures/clustering
• E.g., subtyping autism trajectories
• Reinforcement learning:
• Input varies (online vs offline RL). In offline RL state, action pairs and a reward signal. Goal is to identifying sequence of actions to optimize long term rewards
• E.g., Optimal strategy for sepsis treatment

​

​

Supervised learning: setup

Input: Labeled data

​

​

​

​

​

Supervised learning: setup

Input: Labeled data

​

​

​

​

​

Regression:

Binary classification:

Multiclass-Classification:

Supervised learning: setup

Input: Labeled data

​

​

​

​

​

Goal: learn a mapping h:

Regression:

Binary classification:

Multiclass-Classification:

Picking a hypothesis (mapping)

Define a hypothesis class,

Define evaluation criteria

Define a learning algorithm

Picking a hypothesis (mapping)

Define a hypothesis class,

Define evaluation criteria

Define a learning algorithm

Picking a hypothesis (mapping)

Define a hypothesis class,

Define evaluation criteria

Define a learning algorithm

Age

Income

Space of linear hypotheses

Picking a hypothesis (mapping)

Define a hypothesis class,

Define evaluation criteria

Define a learning algorithm

Age

Education

Age

Education

Space of linear hypotheses

Space of decision-tree based hypotheses

Age

Income

Picking a hypothesis (mapping)

Define a hypothesis class,

Define evaluation criteria

Define a learning algorithm

How good is a prediction?

Picking a hypothesis (mapping)

Define a hypothesis class,

Define evaluation criteria

Define a learning algorithm

How good is a prediction?

Regression problems

Picking a hypothesis (mapping)

Define a hypothesis class,

Define evaluation criteria

Define a learning algorithm

How good is a prediction?

Regression problems

Binary Classification (0/1)

Picking a hypothesis (mapping)

Define a hypothesis class,

Define evaluation criteria

Define a learning algorithm

How good is a prediction?

Regression problems

Binary Classification (0/1)

How good is a hypothesis?

Picking a hypothesis (mapping)

Define a hypothesis class,

Define evaluation criteria

Define a learning algorithm

• A set of steps for the computer to take in order to identify the best
• Usually involves an optimization step�

​

​

What makes an ML model good?

For an unobserved dataset of size m, want low:

​

As a proxy, we will minimize:

​

🡪Empirical risk minimization

​

Warning!

Minimizing training loss is an imperfect proxy. �Will minimizing the training loss give us a good model if we consider the memorization hypothesis?

​

A regression problem

Whiteboard!

Why gradient descent?

• Closed form solution is possible, but expensive (e.g., linear regression)
• Closed form solution is not possible (e.g., logistic regression and neural networks)

Gradient descent: learning rate = 0.05

Gradient descent: learning rate = 0.05

Gradient descent: learning rate = 0.05

Gradient descent: learning rate = 0.05

Gradient descent: learning rate = 0.05

Back to whiteboard…

Gradient descent: Putting it all together

Gradient descent: small learning rate

Gradient descent: large learning rate

Other forms of gradient descent

Stochastic GD

Mini-batch

Adam/Adagrad

Recap

Learning algorithms 🡪 Gradient descent and its sisters

New dataset

New dataset

A different hypothesis class

Polynomial with degree 1

Polynomial with degree 2

Polynomial with degree 3

Polynomial with degree 9

Polynomial with degree 15

Regularization

• Prevents the model from “trying too hard”

Equivalent to:

​

​

Revisiting the hypothesis class

Recall our previous function class definition

​

The regularized version

​

Different norms 🡪 Different hypothesis classes

The bias-variance tradeoff

Thought experiment:

• Sample n points from the full distribution
• Learn a model
• Sample a different set of n points
• Learn a new model
• Repeat…

​

Note: Every model will be slightly different…why?

The bias-variance tradeoff

The bias-variance tradeoff

The bias-variance tradeoff

The bias-variance tradeoff

The bias-variance tradeoff

Low bias, high variance

Large bias, low variance

Low bias, low variance

Holdout method

Observed data

K-fold cross validation, with K = 3

Observed data

Training data

Validation data

K-fold cross validation, with K=3

Observed data

Training data

Validation data

K-fold cross validation, with K=3

Observed data

Training data

Validation data

Training data

K-fold cross validation, with K=3

​

Round 1

�Round 2

�Round 3��

K-fold cross validation, with K=3

​

Round 1

�Round 2

�Round 3��Average

Recap

Hypothesis classes, regularization & bias-variance tradeoff

​

Learning representations

Recall the polynomial regression problem:

​

What if we do not know ?

Feature learning

pixel 1

pixel 2

Input space

Example from J. Wiens

Feature learning

pixel 1

pixel 2

Input space

“muzzle”

“paw”

muzzle

paw

Example from J. Wiens

Neural networks

Chalkboard!

Types of neural networks

• Fully connected neural networks
• Convolutional neural networks
• Recurrent neural networks

Challenges in healthcare settings

• Interpretability
• Bias/fairness
• Causality
• Logistical challenges…