Introduction to machine learning

1. Introduction to machine learning
https://www.springboard.com/blog/machine-learning-engineering/
Introducción al aprendizaje de la máquina
R
N
N
This is not a real cup of coffee
G
A
N
Luc Lesoil
12/03/2020
30/04/2020
14/05/2020
May 20th
2020
DeepMind
Self-driving cars
Cloud optimization
Healthcare
&
Medical images analysis

2. Feel free to
Presentation
I. Supervised learning
II. Unsupervised learning
III. Reinforcement learning
me! :-)

3. Introduction
●
High-level presentation
●
Non-exhaustive list
●
Concrete cases
Math
s
https://github.com/llesoil/ML_example

4. I- Supervised learning
X : Explaining variables
y : Variable to predict, labels known
X : images of numbers y : numbers
3 5 7
4 0
1
Supervised learning = Use X to predict y
Train
the model
Test
the model
3
4
1
0
5
7
3
4
0
5
Learn X→ y Predict ŷ, estimations of y
Compare with real y values
7 1
4
1 =
?
ŷ y

5. Loss function
Minimize the loss function
=
better predictions
Loss function
0
5
2.5
Today’s Tux happiness : 3.2
Prediction 1 : 3
Prediction 2 : 0.5
Compare quality
of predictions?
https://towardsdatascience.com/common-loss-functions-in-mac
hine-learning-46af0ffc4d23
https://algorithmia.com/blog/introduction-to-loss-functions
Tux
MAE
MSE
MAPE
Minkowski
Cross-entropy
Hinge
Examples

6. Supervised learning Qualitative Quantitative
Discrete
int
Continious
float
Nominal
str
Ordinal
str
Colour,
Species
Number
of
occurences
Time,
Temperature
Regression
Classification
Y
Scale
Type of
→ Value
→ Group or category

7. Linear regression
●
Simple
●
Complex dataset
●
Linear relationship
Fit the scatterplot with the red line
is the prediction of x =

8. Decision Tree (CART)
●
Extract rules
●
Simple to parameter
●
Learning unit for many algorithms
X2 < 3.5
X1 < 3.75
→ ŷ = (6.7+6.2)/2

9. Random Forest
●
Bagging→ robustness
●
Metrics
●
Good compromise
= 1 Decision Tree
Democratree !!!
X1
X2
X3
1 7 6
4 3 0
8 1 2
2 4 3

10. Boosting tree
●
Complex dataset
●
Many hyperparameters
●
XGBoost: the algorithm that wins
every competition
+ +
Update the trees based on previous results
AdaBoost
XGBoost

11. Neural networks
●
Simple dataset
●
Many hyperparameters
●
Black box
X1
X2
ŷ
Feedforward neural network

12. Others
●
Quantile/Polynomial/Piecewise regression, Ridge, ElasticNet,
LASSO to select explaining variables
●
Support Vector Machine : SVC or SVR
●
Time series predictions : (S)AR(I)MA, RNN

13. II- Unsupervised learning
●
Clustering
●
Association
●
Anomaly detection
X y

14. Kmeans
●
Simple clustering
●
Fast
●
Few parameters
1
2
3

15. Gaussian Mixture Model
●
Gaussian distribution
●
Estimation of K
●
Scale well - fast

16. K-Nearest neighbors
●
Used in recommendation
systems
●
Supervised
●
“You are the average of the five
people you spend the most time with”
0.6
4.7
1 is the nearest neighbor of 3
0 is the second nearest
2.2

17. Hierarchical clustering
●
Quadratic O(n²)
●
Not designed for big dataset
●
Full description of relationships
4.7
0.6
2.2

18. Hierarchical clustering (2)
●
Quadratic O(n²)
●
Not designed for big dataset
●
Full description of relationships
4.7
0.6
2.2

19. III- Reinforcement Learning
Based on behavioral psychology
Realistic learning
Google - energy consumption -15%

20. Notions
●
Goal
●
(State, Action) → Reward
●
Q-Table
Agent, State
Environment
Reward Action
Policy

21. Map

22. References
●
Mnist dataset, scikit-learn documentation
●
https://github.com/trekhleb/homemade-machine-learning
●
https://internetofbusiness.com/google-using-deepmind-ai-to-reduce-energy-cons
umption-by-30/
●
https://www.slideshare.net/cprakash2011/reinforcement-learning-40052403/5
●
https://brilliant.org/wiki/gaussian-mixture-model/
●
https://perfectial.com/blog/q-learning-applications/
●
MARIQ : https://www.youtube.com/watch?v=CacRZmjDIr4