machine learning introduction - indico · esa unclassified - for official use jose martinez heras |...

ESA UNCLASSIFIED - For Official Use

Machine Learning Introduction

Jose Martinez Heras

16/04/2019

ESA UNCLASSIFIED - For Official Use Jose Martinez Heras | ESOC | 16/04/2019 | Slide 2

About

• José Martínez Heras

• Computer Scientist

• Spanish


Machine Learning Lectures at ESOC

https://github.com/jmartinezheras/2018-MachineLearning-Lectures-ESA



Objectives for this session

• What Machine Learning can / cannot do

• Examples in Space and other domains

• How it works

• Key Concepts & Best Practices

• Understand how some techniques work

• Hands-On session: predict Mars Express thermal power consumption


Artificial Intelligence

Reasoning

Knowledge

Planning

Machine Learning

Natural Language Processing

Perception

Move & manipulate

objects



Reasoning

Knowledge

Planning

Machine Learning


Perception

Move & manipulate

objects


What’s Machine Learning?

• “The field of study that gives computers the ability to learn without being

explicitly programmed” [Artur Samuel, 1959]

• “A computer program is said to learn from experience E with respect to some

task T and some performance measure P, if its performance on T, as measured

by P, improves with experience E” [Tom Mitchell, 1997]

• The science (and art) of programming computers so that they can learn

from data [Aurélien Géron, 2017]


How is Machine Learning different from programming

DataProgram

Perform task

Results DataModel

Perform task

Results

Data

ML

Learning

ModelResults

Requirements Program

Only for some tasks


Which tasks can Machine Learning perform?

- Let’s see some examples

Spam Filter Recommendation Engine



Real Estate

Predict at which price a property will be sold



Advertising

Predict which ads you are more likely to click on



Image Classification



Self-driving cars



Speech Recognition & Synthesis



Language Translation



Playing Games


Some examples in Space Operations

Dr.MUST diagnostics: find the cause of an anomaly

VEX attitude errors because of Aspera payload activation



Novelty Detection: a novel behaviour is often the signature of an

anomaly in the way to happen

OOL

Unusual behaviour

Potential anomaly

XMM Tank1 Ext temperature



[MEX] Predict Thermal Power Consumption: Machine Learning competition



[Integral] Radiation Belts entry / exit times prediction to increase science return



[esa.int] Predict which articles will receive

a high number of views

High: Rosetta, comet, surface, lander, image, crater, mars, stars, galaxy, black holes

Low: ESA, company, technology


Considerations on using Machine Learning

• There must be a pattern in the input –

output relationship (lottery winning

numbers cannot be predicted with ML)

• There must be enough data to

discover this pattern

• It’s difficult to formulate a

mathematical expression (otherwise

we will just use this formula instead)

DataModel

Perform task

Results

Data

ML

Learning

ModelResults


How Machine Learning Learns?

• Different Machine Learning techniques

for different kinds of tasks

• Learning is finding which model’s

parameters represent best the input –

output mapping

DataModel

Perform task

Results

Data

ML

Learning

ModelResults



• Linear Regression example:

• Model: f(x) = mx + b

• Model’s parameters: m, b

• Parameter values: m=1, b=0

• Learning is finding which values of ‘m’

and ‘b’ fit the data best (e.g.

minimizes the prediction error)

DataModel

Perform task

Results

Data

ML

Learning

ModelResults



Find the model parameters values that minimize the error



m = 0.1014b = 1.2258

Find the model parameters values that minimize the error


How Machine Learning Learns? – some models

Polynomials: ax + bx2 + … + c

Support Vector Machines (SVM): Vectors

Decision trees: which nodes, which decisions


How Machine Learning Learns? – some models

Neural Networks / Deep Learning: weights


Types of Learning

Depending on the level of supervision …

• Supervised

• Unsupervised

• Semi-supervised

• Reinforcement Learning


Types of Learning

• Supervised

• Supervision: we can tell for every case what the correct answer was

• Example: predict the thermal power consumption

FocusPredict the future


Types of Learning

• Unsupervised

• Supervision: there is no right answer, we are looking for insights

• Example: market basket analysis for supermarkets

FocusUnderstand the past


Types of Learning

• Semi-supervised

• Supervision: we can tell the correct output for a limited number of cases

• Example: characterize what a particle impact looks like in TM

Sentinel-1A: particle impact on August 23th 2016

FocusUnderstand the past


Types of Learning

• Reinforcement Learning

• Supervision: we only know the final outcome, but not intermediate steps

• Example: playing Go

FocusFind which is the next action most likely to lead to the desired outcome


Types of Learning

• The type of learning with most industrial applications is

Supervised Learning

(Predictive Analytics)

• Depending of what kind of data is predicted we can talk about:

• Regression

• Classification


Regression

• Predict real numbers


Classification

• Predict which option out of a limited set of possibilities

Spam Filter What’s in the picture


Machine Learning Workflow

Problem Understanding




Evaluation Criteria

• How will we measure how good the model is performing?

• Do we know already at what point it would be enough?




Evaluation Criteria

Baseline

• How the current approach is performing against the evaluation criteria? Define a simple baseline if there is none (e.g. mean value)

• This will allow us to quantify how much Machine Learning helps and if it is worthwhile compared to simpler solutions




Evaluation Criteria

Baseline

Prepare Data

• “Enough” data in the sense that it’s representative of the behaviour the model needs to learn

• Features: data transformations that encode your knowledge




Evaluation Criteria

Baseline

Prepare Data Train model

• Use data / features to tune the parameters that optimize the evaluation criteria (e.g. minimise error)




Evaluation Criteria

Baseline

Prepare Data Train model Error Analysis

More complex model

Simpler model

More data / features

Provides better problem understanding

• Understand what the model is doing: where is it right / wrong




Evaluation Criteria

Baseline


More complex model

Simpler model


Provides better problem understanding

ProductionTrained Model

Ready to use

Better baseline

• When we are happy with the error, we can use the trained model: new data features computation model results

• We now have a better baseline




Evaluation Criteria

Baseline


More complex model

Simpler model


Better problem understanding


Ready to use

Better baseline


Evaluation Criteria

Typically a way to measure error.

Example:

𝑀𝑆𝐸 =1

𝑛

𝑖=1

𝑛

𝑌𝑖 − 𝑌𝑖2

Mean Squared Error

* We will see more evaluation criteria in other sessions


Train Machine Learning Models

Let’s train 2 models to learn from this data

𝑦 = 0.1 ∗ 𝑥 + 1.25 + 0.2 ∗ 𝐺𝑎𝑢𝑠𝑠𝑖𝑎𝑛𝑁𝑜𝑖𝑠𝑒


Train Machine Learning Models

Mean Squared Error: 0.0372 Mean Squared Error: 0.0286

Linear Regression Polynomial Regression (degree 7)


Generalization

Generalization: the ability of a model to perform well on new data


Generalization

How to tell if a Machine Learning model is generalizing well when we cannot

visualize what it is doing?

?

2-dimensional data 3-dimensional data n-dimensional data


Generalization

Split data in 2 sets

• Training Set: for training the model

• Testing Set: for evaluating the generalization of the model

* Validation set and cross-validation will be explained later


Generalization

Data Training Data + Testing Data

The Machine Learning model now

learns from training data only


Generalization – Error Analysis

Evaluate Performance on Training Data & Test Data

MSE Train: 0.043, MSE Test: 0.0238 MSE Train: 0.043, MSE Test: 85.113

Over-fitting


Generalization – Error Analysis - Guidelines

Train Error Test Error What to do?

Low High • Need a simpler model• Need more data (more data samples)

* High / Low compared with the error you are willing to accept

Over-fitting





High High • Need more data• Difficult to learn f(x, z) with only x. Get also z• Get more data samples

• Additional features (e.g. 1

𝑥2)

• Need a more complex model


Over-fitting

Under-fitting







𝑥2)


High Low Unusual: it could mean that the test data is too similar to the train data. Get more test data.


Over-fitting

Under-fitting







𝑥2)


High Low Unusual: it could mean that the test data is too similar to the train data. Get more test data.

Low Low You’re done! – congratulations


Over-fitting

Under-fitting




Evaluation Criteria

Baseline


More complex model

Simpler model


Better problem understanding


Ready to use

Better baseline

Effort


Let’s talk about Random Forests

• Decision Trees

• Ensembles

• Random Forests


Random Forests

[MEX] Predict Thermal Power Consumption


Random Forests


[Integral] Radiation Belts

entry / exit times prediction

to increase science return


Random Forests

Random Forest = Ensemble of Decision Trees

1. Decision Trees

2. Ensembles

3. Random Forests


Decision Trees – remember Iris example?

Iris setosa Iris versicolor Iris virginica

Knowing the sepal and petal length and width, which flower is it?

Pictures from Wikipedia contributors, "Iris flower data set," Wikipedia, The Free Encyclopedia, https://en.wikipedia.org/w/index.php?title=Iris_flower_data_set&oldid=824486644

(accessed March 5, 2018).


Decision Trees

It also can provide probabilities

𝑝𝑠𝑒𝑡𝑜𝑠𝑎 = 1.00

Gini is a measurement of purity

𝑝𝑣𝑒𝑟𝑠𝑖𝑐𝑜𝑙𝑜𝑟 = 0.91 𝑝𝑣𝑖𝑟𝑔𝑖𝑛𝑖𝑐𝑎 = 0.98


Decision Trees

𝐺𝑖𝑛𝑖 = 1 −

𝑘=1

𝑛

𝑝𝑐2

𝐺𝑖𝑛𝑖 = 1 −0

54

2

+49

54

2

+5

54

2

= 0.168

Gini is a measurement of purity

𝑝𝑐= probability of each class


How do we learn to build Decision Trees?

Greedy optimization that minimizes

𝐽 𝑓, 𝑡𝑓 =𝑚𝑙𝑒𝑓𝑡

𝑚𝐺𝑖𝑛𝑖𝑙𝑒𝑓𝑡 +

𝑚𝑟𝑖𝑔ℎ𝑡

𝑚𝐺𝑖𝑛𝑖𝑟𝑖𝑔ℎ𝑡

𝑓 = which feature

𝑡𝑓 = which threshold for feature f

𝑚 = number of samples


Features Importance

Features Importance

sepal length (cm): 0.0sepal width (cm): 0.0petal length (cm): 0.562petal width (cm): 0.438


Decision Trees - Regression

𝑦 = 𝑥2 + 0.2 ∙ 𝐺𝑎𝑢𝑠𝑠𝑖𝑎𝑛𝑁𝑜𝑖𝑠𝑒



𝑦 = 𝑥2 + 0.2 ∙ 𝐺𝑎𝑢𝑠𝑠𝑖𝑎𝑛𝑁𝑜𝑖𝑠𝑒 𝐽 𝑓, 𝑡𝑓 =𝑚𝑙𝑒𝑓𝑡

𝑚𝑀𝑆𝐸𝑙𝑒𝑓𝑡 +


𝑚𝑀𝑆𝐸𝑟𝑖𝑔ℎ𝑡


Mean Squared Error

𝑀𝑆𝐸 =1

𝑚

𝑖=1

𝑚

𝑌𝑖 − 𝑌𝑖2

Mean Squared Error



𝑦 = 𝑥2 + 0.2 ∙ 𝐺𝑎𝑢𝑠𝑠𝑖𝑎𝑛𝑁𝑜𝑖𝑠𝑒 𝐽 𝑓, 𝑡𝑓 =𝑚𝑙𝑒𝑓𝑡

𝑚𝑀𝑆𝐸𝑙𝑒𝑓𝑡 +


𝑚𝑀𝑆𝐸𝑟𝑖𝑔ℎ𝑡


Regularization in Decision Trees

overfitting tree

For this particular example. Require thateach node has ≥ 5% of the samples


Regularization in Decision Trees

• Maximum depth

• Maximum number of leaf nodes

• Minimum number of samples at leaf nodes

• Minimum samples at node must have before it can be split


Decision Trees don’t make assumptions

This is good because it’s not limited by a line, polynomial or any other

predefined model

𝑦 = 𝑥2 + 0.2 ∙ 𝐺𝑎𝑢𝑠𝑠𝑖𝑎𝑛𝑁𝑜𝑖𝑠𝑒 𝑦 = sin(1.5 ∙ 𝑥) + 0.2 ∙ 𝐺𝑎𝑢𝑠𝑠𝑖𝑎𝑛𝑁𝑜𝑖𝑠𝑒


Decision Trees don’t make assumptions

But it only works well when trained with representative data

𝑦 = 𝑥2 + 0.2 ∙ 𝐺𝑎𝑢𝑠𝑠𝑖𝑎𝑛𝑁𝑜𝑖𝑠𝑒 𝑦 = sin(1.5 ∙ 𝑥) + 0.2 ∙ 𝐺𝑎𝑢𝑠𝑠𝑖𝑎𝑛𝑁𝑜𝑖𝑠𝑒


Ensembles

• Ensemble = group of Machine Learning models

• Each model in the group produces a different prediction

• Predictions are combined to give a final prediction

• How predictions are combined?

• voting

• bagging

• boosting

• stacking


Ensembles - Voting

Logistic Regression

Support Vector Machine

Decision Tree

New sample

1

0

1

1

It works better than individual modelsAs long as models are independent, their errors compensate each other

Already trained models


Ensembles - Bagging

• Voting of several instances of the same model

• majority voting from several trees

• However

• all tress would be so similar that their errors will not compensate

• Let’s make Decision Trees different so that we can combine them


Ensembles - Bagging

Training Data

Random sampling

Random Sampling

Random Sampling

model

model

model

1

0

1

1

Training the models Using the models

Train decision models with different data / features

New sample


Random Forest: Bagging Ensemble of Trees

Training Data

Random sampling

Random Sampling

Random Sampling

Decision Tree

Decision Tree

Decision Tree

Train decision models with different data / features

1

0

1

1

Training the models Using the modelsNew sample


Random Forests

Random Forest are good for generalization

Random Forest provide a similar training error (same bias) as an individual treeRandom Forest provide lower testing error (lower variance) as an individual tree


How results are combined for Random Forest?

Classification

• Soft-voting

• Combine the probabilities:

highly confident predictions

get higher weight

Regression

• Mean value

You can also implement your own combination of predictions


How results are combined for Random Forest?

Custom combinations of the Decision Trees Ensembles


Decision Trees or Random Forest?

Decision Tree

• You are interested in a

white-box model

• Understanding is more

important than predicting

Random Forest

• You don’t mind using a

black-box model

• Predicting (generalizing

well on new data) is

more important than

understanding

Tree + Random Forest

• Decision Tree for

understanding

• Random Forest for

predicting

• Use Tree as surrogate of

the Random Forest


Surrogate Models

Black Box(e.g. Random Forest)

Results

FeaturesPredicted Results


Surrogate Models

Black Box(e.g. Random Forest)

Results

FeaturesPredicted Results

White Box(e.g. Decision Tree)

Predicted Results (II)


Random Forests

Random Forests gives you forecasting super-powers

But only if new data is similar to training data

It’s not as bad as I pictured it

Because, in practice, there are many more features


Let’s recap

Today we have discussed about:

• What is machine learning

• Some applications in industry & space

• What is the ‘Learning’ in Machine Learning

• Types of machine learning

• Regression / Classification

• Machine Learning workflow

• Generalization

• Decision Trees

• Ensembles

• Random Forests

• Surrogates models for Explainable AI


Thank you

LinkedIn: https://www.linkedin.com/in/josemartinezheras/


https://www.linkedin.com/in/josemartinezheras/


ESA UNCLASSIFIED - For Official Use

Thank you

LinkedIn: https://www.linkedin.com/in/josemartinezheras/

Jose Martinez Heras

https://www.linkedin.com/in/josemartinezheras/

machine learning introduction - indico · esa unclassified - for official use jose martinez heras |...

Documents