Introduction to Machine Learning

Sean Davis, MD, PhD

Friday, September 20, 2024

Outline

1. Relationship between AI and ML
2. Types of Machine Learning
3. A little about data
4. Supervised Learning
5. Unsupervised Learning

Artificial Intelligence Overview

Artificial Intelligence (AI) is the simulation of human intelligence processes by machines, especially computer systems. These processes include learning (the acquisition of information and rules for using the information), reasoning (using rules to reach approximate or definite conclusions), and self-correction.

An Ontology of AI

What is Machine Learning?

The study of computer algorithms that improve automatically through experience. It is seen as a subset of artificial intelligence. Machine learning systems give the computer the ability to learn without being explicitly programmed rules.

What is Deep Learning?

Machine learning algorithms that are inspired by the structure and function of the brain. Deep learning is a subset of machine learning in artificial intelligence that has networks capable of learning unsupervised from data that is often unstructured (i.e., text or images).

AI vs. ML vs. Deep Learning

Artificial
Intelligence

Machine
Learning

Deep
Learning

AI vs. ML vs. Deep Learning

Artificial
Intelligence

Machine
Learning

Deep
Learning

Side trip: Types and Characteristics of Data

An ontology of Data

One way to think about different classes of data.

Structured Data vs Unstructured Data

	Structured data	Unstructured data
What is it?	Data that fits in a predefined data model or schema.	Data without an underlying model to discern attributes.
Basic example	An Excel table.	A collection of video files.
Best for	A set of predefined observations or characteristics (columns) on a collection of things (rows)	An associated collection of data, objects, or files where the attributes change or are unknown.
Common formats	CSV, TSV, Excel, SQL databases.	Images, audio, video, text.

Tabular Data

• Rows and columns
• Each column has a specific data type
• Each row represents an observation on a subject (e.g., patient, sample)

Terminology:

• Feature: A column in the dataset
• Target: The variable you are trying to predict
• Observation: A row in the dataset
• Dimension: The number of features in the dataset (sometimes represented by \(p\))
• Sample: A single row in the dataset (usually represented by \(n\))

Classes of Machine Learning

Broad classes of machine learning.

Types of Machine Learning

1. Supervised Learning
   • Classification
   • Regression
2. Unsupervised Learning
   • Clustering
   • Dimensionality Reduction
3. Reinforcement Learning

Reinforecement Learning

• Learning through interaction: Unlike supervised learning where you have labeled data, RL agents learn by interacting with an environment.
• Delayed rewards: The agent doesn’t receive immediate feedback about the optimal action, but rather must learn which actions lead to better cumulative rewards over time.
• Exploration vs. exploitation: The agent must balance exploring new actions versus exploiting known good actions.
• Sequential decision making: RL deals with sequences of decisions rather than one-time predictions.

Examples

• Game playing (e.g., AlphaGo, OpenAI Five)
• Robotics control

• Autonomous driving
• Trading strategies

A map of machine learning approaches

Supervised Learning

Supervised learning builds a model on which we know the TRUE result; this is sometimes called “labeled data”.

The model learns to map input data to the correct output by adjusting its parameters to minimize the error between the predicted output and the true output.

Supervised Learning

• Given a dataset of input-output pairs: \(\{(x_1, y_1), (x_2, y_2), \ldots, (x_n, y_n)\}\)

• Learn a function \(f\) with parameters \(\theta\) that maps inputs \(x\) to outputs \(y\) \[y = f(x; \theta) + \epsilon\]

• Objective: Choose parameters \(\theta\) to minimize the prediction error \(\epsilon\)

Supervised Learning

When \(y\) is a continuous variable (e.g., house prices), it’s a regression problem.

When \(y\) is a discrete variable (e.g., spam or not spam), it’s a classification problem.

It is not a concidence that the same terms are used in statistics! Many machine learning algorithms are based on statistical principles or are generalizations of statistical methods.

Classification

Regression

Not all regression is linear….

Classification Example

The classic iris

Classification Example

Sepal.Length	Sepal.Width	Petal.Length	Petal.Width	Species
6.2	2.8	4.8	1.8	virginica
4.8	3.4	1.6	0.2	setosa
5.0	3.4	1.5	0.2	setosa
5.9	3.2	4.8	1.8	versicolor
5.8	4.0	1.2	0.2	setosa
5.8	2.7	3.9	1.2	versicolor
6.7	3.1	4.7	1.5	versicolor
5.6	2.8	4.9	2.0	virginica
6.3	2.8	5.1	1.5	virginica
6.3	2.5	5.0	1.9	virginica
4.6	3.1	1.5	0.2	setosa
5.5	2.4	3.8	1.1	versicolor
6.5	3.0	5.2	2.0	virginica
5.6	3.0	4.5	1.5	versicolor
7.3	2.9	6.3	1.8	virginica
6.1	3.0	4.9	1.8	virginica
4.3	3.0	1.1	0.1	setosa
6.4	3.2	5.3	2.3	virginica
4.8	3.0	1.4	0.3	setosa
6.1	2.9	4.7	1.4	versicolor

Classification Example

Training and Testing

• Training Data: Used to teach the model. It consists of input-output pairs for a supervised learning task.
• Testing Data: Used to evaluate the model’s performance. It should be separate from the training data, but come from the same distribution or population.
• Validation Data: Used for tuning hyperparameters. Hyperparameters are settings that control the learning process (e.g., k in k-NN, learning rate in neural networks).

Generalizability

• The ability of a model to perform well on unseen data

This table shows the relationship between training error, testing error, and overfitting, good fit, and underfitting.

Training Error	Testing Error	Model Fit
Low	High	Overfit
Low	Low	Good Fit
High	High	Underfit (poor performance)

Bias and Variance

• Bias: Error from incorrect assumptions in the learning algorithm
• Variance: Error from sensitivity to small fluctuations in the training set
• Bias-Variance Tradeoff: The conflict in trying to simultaneously minimize these two sources of error

Classification Example

Classification Example

Classification Example

Regression Example

Call:
lm(formula = y ~ x)

Residuals:
    Min      1Q  Median      3Q     Max 
-49.071 -11.047  -0.692  12.970  41.897 

Coefficients:
            Estimate Std. Error t value Pr(>|t|)    
(Intercept) -0.72808    3.68575  -0.198    0.844    
x            2.05022    0.06336  32.356   <2e-16 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 18.29 on 98 degrees of freedom
Multiple R-squared:  0.9144,    Adjusted R-squared:  0.9135 
F-statistic:  1047 on 1 and 98 DF,  p-value: < 2.2e-16

Regression Example

Call:
lm(formula = y ~ x)

Residuals:
    Min      1Q  Median      3Q     Max 
-98.143 -22.095  -1.385  25.940  83.795 

Coefficients:
            Estimate Std. Error t value Pr(>|t|)    
(Intercept)  -1.4562     7.3715  -0.198    0.844    
x             2.1004     0.1267  16.574   <2e-16 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 36.58 on 98 degrees of freedom
Multiple R-squared:  0.7371,    Adjusted R-squared:  0.7344 
F-statistic: 274.7 on 1 and 98 DF,  p-value: < 2.2e-16

Regression Example

Call:
lm(formula = y ~ x)

Residuals:
     Min       1Q   Median       3Q      Max 
-171.749  -38.665   -2.423   45.395  146.641 

Coefficients:
            Estimate Std. Error t value Pr(>|t|)    
(Intercept)  -2.5483    12.9001  -0.198    0.844    
x             2.1758     0.2218   9.811 3.12e-16 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 64.02 on 98 degrees of freedom
Multiple R-squared:  0.4955,    Adjusted R-squared:  0.4904 
F-statistic: 96.25 on 1 and 98 DF,  p-value: 3.116e-16

More Complex Regression Example

age	sex	bmi	children	smoker	region	charges	obese
19	female	27.900	0	yes	southwest	16884.924	not obese
18	male	33.770	1	no	southeast	1725.552	obese
28	male	33.000	3	no	southeast	4449.462	obese
33	male	22.705	0	no	northwest	21984.471	not obese
32	male	28.880	0	no	northwest	3866.855	not obese
31	female	25.740	0	no	southeast	3756.622	not obese
46	female	33.440	1	no	southeast	8240.590	obese
37	female	27.740	3	no	northwest	7281.506	not obese
37	male	29.830	2	no	northeast	6406.411	not obese
60	female	25.840	0	no	northwest	28923.137	not obese

More Complex Regression Example

What do you think about the relationship between age and insurance charges?

More Complex Regression Example

More Complex Regression Example

Call:
lm(formula = charges ~ age + smoker + sex + obese, data = insurance)

Residuals:
     Min       1Q   Median       3Q      Max 
-13821.5  -3647.0   -225.8   1491.3  26884.1 

Coefficients:
            Estimate Std. Error t value Pr(>|t|)    
(Intercept) -4051.16     542.42  -7.469 1.46e-13 ***
age           261.73      11.82  22.149  < 2e-16 ***
smokeryes   23861.89     410.79  58.087  < 2e-16 ***
sexmale      -114.30     331.89  -0.344    0.731    
obeseobese   4234.44     332.58  12.732  < 2e-16 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 6044 on 1333 degrees of freedom
Multiple R-squared:  0.7516,    Adjusted R-squared:  0.7509 
F-statistic:  1008 on 4 and 1333 DF,  p-value: < 2.2e-16

Supervised Machine Learning Models

Algorithms for Supervised Learning

Call:
lm(formula = y ~ x)

Residuals:
     Min       1Q   Median       3Q      Max 
-24.5356  -5.5236  -0.3462   6.4850  20.9487 

Coefficients:
            Estimate Std. Error t value Pr(>|t|)    
(Intercept) -0.36404    1.84287  -0.198    0.844    
x            2.02511    0.03168  63.920   <2e-16 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 9.145 on 98 degrees of freedom
Multiple R-squared:  0.9766,    Adjusted R-squared:  0.9763 
F-statistic:  4086 on 1 and 98 DF,  p-value: < 2.2e-16

Algorithms for Supervised Learning

Anscombe’s Quartet

Algorithms for Supervised Learning

K-nearest neighbor (kNN) algorithm.

Algorithms for Supervised Learning

Classification and Regression Trees (CART).

Algorithms for Supervised Learning

Random Forests.

Algorithms for Supervised Learning

Deep learning.

Applying Supervised Learning Algorithms

The mlr3 ecosystem in R.

Key Concepts in Supervised Learning

1. Feature Engineering
2. Model Selection
3. Hyperparameter Tuning
4. Ensemble Methods
5. Cross-Validation
6. Evaluation Metrics

Briefly explain each concept: 1. Feature Engineering: Creating new features or transforming existing ones to improve model performance 2. Model Selection: Choosing the appropriate algorithm for your problem 3. Hyperparameter Tuning: Optimizing model parameters that are not learned from data 4. Ensemble Methods: Combining multiple models to improve performance 5. Cross-Validation: Technique for assessing how the model will generalize to an independent dataset 6. Evaluation Metrics: Different ways to measure model performance (accuracy, precision, recall, F1-score, ROC curve, etc.)

Challenges in Machine Learning

1. Data Quality and Quantity
2. Interpretability vs Performance
3. Ethical Considerations and Bias
4. Computational Resources
5. Model Deployment and Maintenance

Discuss each challenge: 1. The importance of good, representative data and the challenges of data collection and cleaning 2. The tradeoff between complex, high-performing models and simpler, more interpretable ones 3. The risk of perpetuating or amplifying societal biases through ML models 4. The need for significant computational power, especially for deep learning models 5. The challenges of deploying models in production environments and keeping them updated

Unsupervised Learning

Unsupervised learning.

Clustering

Gene expression measurements.

Classes of unsupervised learning algorithms

Clustering.

Dimensionality reduction.

Clustering

Thinking about similarities and differences

Clustering

Algorithm for hierarchical clustering

Clustering

Algorithm for hierarchical clustering

Clustering

Algorithm for hierarchical clustering

Dimensionality Reduction

Schematic PCA.

Dimensionality Reduction

Using dimensionality reduction to explore 22,000 dimensions of gene expression data on 280 samples.

Future Directions in Machine Learning

• AutoML: Automating the ML pipeline
• Federated Learning: Training models on decentralized data
• Explainable AI: Making black-box models more interpretable
• Quantum Machine Learning: Leveraging quantum computing for ML
• Continual Learning: Adapting to new data without forgetting old patterns

Conclusion

• Supervised learning: Predicting outcomes based on labeled data
  • Classification: Assigning labels to data
  • Regression: Predicting continuous values
• Unsupervised learning: Finding patterns in unlabeled data
  • Clustering: Grouping similar data points
  • Dimensionality reduction: Simplifying complex data by reducing dimensions
• Reinforcement learning: Learning through interaction with an environment

Resources

• Recent reviews
  • Machine Learning in Genomic Medicine: A Review of Computational Problems and Data Sets
  • Opportunities And Obstacles For Deep Learning In Biology And Medicine
• Hands-on Tutorials
  • https://seandavi.github.io/RBiocBook
  • Many online courses and tutorials
• Blogs and online materials
  • https://blog.recast.ai/machine-learning-algorithms/
  • https://www.youtube.com/channel/UCWN3xxRkmTPmbKwht9FuE5A
  • https://www.youtube.com/c/joshstarmer
• Machine Learning in R
  • https://www.datacamp.com/community/tutorials/machine-learning-in-r
  • https://daviddalpiaz.github.io/r4sl/

Questions?