Overview of Machine Learning and Pattern Recognition

Outline

Pattern Recognition and Machine Learning: de nitions and

main concept
The classi cation paradig
Evaluating classi cation/prediction performanc
Feature Reductio
Supervised/Unsupervised learning (overview and examples of
approaches
Neural networks and Deep learning (overview and examples)

Bioinformatics – Overview of Machine Learning and Pattern Recognition 2

fi
)

fi
n

fi
e

1. Introduction
 Data provided by sequencing
• Genes/TFs/miRNAs/short or long non coding RNAs Expression
• Mutations (single base or group of bases)
• CNVs
• miRNAs and isomiRNAs
• miRNA/non-coding RNAs targets
• Binding sites for TFs, regulative proteins, regulative non coding RNAs
• Gene fusions
• Information about fusions structural mechanisms
• Epigenetics, conserved DNA regions
• Virus DNA/RNA damages
• Regulatory networks
• Proteins expression, networks etc.
Bioinformatics – Overview of Machine Learning and Pattern Recognition
 Machine learning and sequencing

• These data can be used as input of classifiers in order to

– Discovery drivers, features, very important molecular players
or very relevant molecular interactions
– Build living systems models for simulations
– Build predictive models for pathology diagnosis, prognosis and
therapy

Bioinformatics – Overview of Machine Learning and Pattern Recognition

 A doll inside a doll
Computer Science

Artificial Intelligence

Machine Learning

Pattern Recognition

Bioinformatics – Overview of Machine Learning and Pattern Recognition

 Artificial Intelligence
• “Intelligence” (i.e. cognitive abilities typically associated to human mind, such as
learning, making decisions, problem solving, etc.) exibited by a machin
• Branch of computer science aimed at developing arti cial agents able t
– Perceive the environment (i.e. receive data from the external world
– Take actions maximising the chance of success at some goal

Bioinformatics – Overview of Machine Learning and Pattern Recognition

fi
)

 Machine Learning
• Subfield of artificial intelligence that gives computers the
ability to learn how to solve a specific problem without being
explicitly programmed for i
• Learning = Deriving knowledge from experimental data

Bioinformatics – Overview of Machine Learning and Pattern Recognition

t

 Pattern Recognition

• Most of the times it is used as a synonym of machine learnin

• More precisely, it is a branch of machine learning that focuses on
the recognition of patterns and regularities in dat
• Pattern = Any kind of discernible regularity in a set of data, whose elements
repeat in a predictable manner

Examples
– Specific patterns of DNA sequences (genes, protein coding regions, promoters,
etc.) uncover functional aspects of cell
– Specific visual patterns of biological images allow to identify cell types,
unveil cancer, etc.
Bioinformatics – Overview of Machine Learning and Pattern Recognition
:

 What is Machine Learning?

• Machine Learning: study of algorithms that improve their
performance at some task with experience

• System that can continuously self-improve and thereby offer

increased efficiency and effectiveness by learning from experience,
analytical observation, and other means.

• It’s a change of paradigm: Knowledge-based vs. Learning

Systems
Bioinformatics – Overview of Machine Learning and Pattern Recognition
 A change of paradigm
Traditional Programming
Data
Computer Output
Program

Machine Learning
Data Model/
Output Computer
Program
Bioinformatics – Overview of Machine Learning and Pattern Recognition
 Knowledge-based vs. Learning Systems
• Knowledge-based Systems: Acquisition and modeling of common-sense
knowledge and expert knowledge
⇒ limited to given knowledge base and rule se
⇒ Inference: Deduction generates no new knowledge but makes implicitly
given knowledge explici
⇒ Top-Down: from rules to fact

• Learning Systems: Extraction of knowledge and rules from examples/

experience
⇒ Learning as inductive proces
⇒ Bottom-Up: from facts to rules
Bioinformatics – Overview of Machine Learning and Pattern Recognition
t

 Why “Learn”?
• Learning is necessary when
• Human expertise does not exis
• Humans are unable to explain their expertis
• Human expertise exists, but it is unreliable (e.g. result may be affecte
❖ by subjectivity
• Human expertise exists, but it is unfeasible (e.g. too many data to process, or too costly
• Solution may change in tim
• Solution may need to be adapted to particular case
• Generally, data is cheap and abundant, while expertise is expensive and scarce

Bioinformatics – Overview of Machine Learning and Pattern Recognition

)

 Growth of Machine Learning

Machine learning is the preferred approach to
Speech recognition, Natural language processing
Computer vision
Robot contro
Medical outcomes analysi
Computational biology

This trend is acceleratin

Improved machine learning algorithms
Improved data capture, networking, faster computer
Demand for self-customization to use
It turns out to be difficult to extract knowledge from human experts ! failure
of expert systems in the–1980’s.
Bioinformatics Overview of Machine Learning and Pattern Recognition
l

 Types of machine learning

Supervised Unsupervised
Learning Learning

Reinforcement
Learning
Bioinformatics – Overview of Machine Learning and Pattern Recognition
 Supervised learning

Bioinformatics – Overview of Machine Learning and Pattern Recognition

 Two types of supervised learning

• Classi ca on
Label: Label:
1 2

• Regression

Bioinformatics – Overview of Machine Learning and Pattern Recognition

fi
ti
 Classification
A subcategory of supervised learning where the goal is to predict the
categorical class labels of new instances, based on a training set of past
observations
Class labels: discrete, unordered values representing the group memberships
of the instances
Traning set: A set of instances with known class labels, based on which
the classifier learns a classification strategy than can be applied to new
unlabelled instances

Binary classification ! two classes

Multi-class classification ! more than two classes
Bioinformatics – Overview of Machine Learning and Pattern Recognition
 Binary classification
Learnt classification model:
DECISION BOUNDARY

+: Positive clas
O: Negative class

Bioinformatics – Overview of Machine Learning and Pattern Recognition

s

 Regression

• Differently from classification, that assigns categorical class labels

to the instances, the aim of regression analysis is prediction of
continuous outcomes
• Given a number of predictor (explanatory) variables and a
continuous response variable (outcome), regression analysis tries to
find a mathematical relationship between those variables, which can
be used to predict the outcome with a resonable level of
approximation.

Bioinformatics – Overview of Machine Learning and Pattern Recognition

 Example: linear regression

x: predictor variable
y: outcome (response variable)

Linear regression: nd the line that

minimizes the distance (typically, the
mean squared distance) between all
the sample points and the fitted line

Bioinformatics – Overview of Machine Learning and Pattern Recognition

fi
 Reinforcement learning
• The goal is to develop a system (agent) that improves its
performance based on interactions with the environment
• The agent takes an action on each time step and receives a reward
value, which is a measure of how good the action is towards the
goal to be achieved

Bioinformatics – Overview of Machine Learning and Pattern Recognition

 Reinforcement learning
• Through the interaction with the environment, the agent learns a
series of actions that maximizes this reward, typically via an
exploratory trial-and-error approach.

• Most popular example: chess engine.

The agent decides upon a set of

possible moves based on the current
state of the board (environment) and
the reward (win or lose at the end of
the game)
 Chess example (1)
• Let's say you start with a chess board set up for the start of a
game. Each player has 16 pieces. Let's say that white starts.
White has 20 possible moves
– The white player can move any pawn forward one or two positions
– The white player can move either knight in two different ways

• The white player chooses one of those 20 moves and makes it

• For the black player, the options are the same: 20 possible
moves. So black chooses a move.
Bioinformatics – Overview of Machine Learning and Pattern Recognition
:

 Chess example (2)

• In a world of "all possible moves," the program makes a big tree for all of
those moves: The total number of board positions is 10120

• No computer is ever going to calculate the entire tree. What a chess

computer tries to do is generate 5 or 10 or 20 moves into the future (e.g. a
ve-level tree contains 3,200,000 board positions

• Once it generates the tree, then the computer needs to "evaluate the board
positions." That is, the computer has to look at the pieces on the board and
decide whether that arrangement of pieces is "good" or "bad"

• The way it does this is by using an evaluation function. The simplest

possible function might just count the number of pieces each side has,
and– then
Bioinformatics compute
Overview the difference.
of Machine Learning and Pattern Recognition
fi
)

 Chess example (3) –Evaluation function

• The previous formula can be more complicated by applying a weight

to each type of piece

• The evaluation function becomes more and more complicated by

adding things like:
board position, control of the center, vulnerability of the king to check,
vulnerability of the opponent's queen, and tons of other parameters

• No matter how complicated the function gets, however, it is

condensed down to a single number that represents the "goodness" of
that board position.
Bioinformatics – Overview of Machine Learning and Pattern Recognition
.

 Reinforcement Learning

• Differently from supervised learning, it does not build a system

based on a training set with apriori known outcomes. It just works
towards the maximization of the cumulative reward

• The sub-optimal actions (i.e. actions with negative reward) are not
necessarily to be avoided: exploration–exploitation tradeof
• Main application area
– game theor
– robotics
y

 Unsupervised Learning

• Deals with unlabeled data or data of unknown structure

• The algorithm explores the inherent structure of the data, without the
guidance of a labelled training set or reward function. Typical
applications

– Clustering (a.k.a. unsupervised classification)

– Dimensionality reduction
:

 Clustering
The instances are partitioned into a number of classes
(clusters) based on the
– Maximization of similarity of instances of the same cluste
– Minimization of similarity of instances of different cluters

Examples of applications in bioinformatics:

– Exploration of recurrent sequence motif
– Discrimination of tissue types in biological image
– Automatic summarization
s

 Dimensionality reduction
• Data of high dimensionality, each observation comes with a high
number of measurements, can present a challeng
– limited storage spac
– computational performance of machine learning algorithm

• Dimensionality reduction projects the input instances into a new

lower-dimensional space, in order t
– Remove nois
– Retain the most relevant informatio

• Most of the times, it is a pre-processing step

e

 In a nutshell…
SUPERVISED REINFORCEMENT UNSUPERVISED

• The algorithm presents • Builds an algorithm

• Builds a classifier based on
a state and takes an based on input data
input and outp ut data
action based on the • That algorithm is tested with
• Classifier is trained input data a test set of data (in which
with a training set of • The action is the algorithm creates the
data rewarded or punished classifier)

• Classifier is tested with a • The algorithm learns from

test set of data the reward/punishmen t • Deployment if the
• Deployment if the and updates itself, this classifier is satisfactory
output is satisfactory continues
 Resources: Journals
Journal of Machine Learning Research
[Link]
Machine Learning
IEEE Transactions on Neural Networks
IEEE Transactions on Pattern Analysis and
Machine Intelligence
Annals of Statistics
Journal of the American Statistical Association
Pattern Recognition
…
 Resources: Conferences
International Conference on Machine Learning (ICML)
European Conference on Machine Learning (ECML
Neural Information Processing Systems (NIPS)
Computational Learning
International Joint Conference on Arti cial Intelligence
(IJCAI)
ACM SIGKDD Conference on Knowledge Discovery and
Data Mining (KDD
IEEE Int. Conf. on Data Mining (ICDM)
…
)

fi
)

 Resources: Software
Matlab: Easy to learn, very powerful and comprehensive, you can find nearly every high-
level function you need and put them together to satisfy your needs. The speed is optimized
if you use vectorized and matrix computation.

Weka (Java): Minimal programming skill required if you use GUI. Also Java is a powerful
language for wri ng your own algorithms. Details (like input, output) can be handled easily.

R: Free and powerful. More powerful than Matlab if you are doing statistical modeling,
but inferior in its general toolbox.

Python: There are libraries like sklearn, numpy, pandas, and scipy making python
competent to do all kinds of scientific computing with its rich and powerful language
features.
ti