ADVANCED
MACHINE LEARNING
Slide 2
• Knowledge representation
• Definition of machine learning
• Sample digitizing
• Learning types
• Learning paradigma
• Learning strategies
Knowledge Representation
In this part we will mainly focus on
digitizing the information
All the events in the world are needed to
be described as some digital or analog
numbers
Examples
For the currency prediction:
The Inputs:
I1: The currency value for this term
I2: Inflation rate
I3: Consumer index
I4: Stock index
I5: Periodically exportation inrease/decrease
rate
The Output: O: The currency value for the next
term
Sexuality: For male “1” otherwise “0”.
If the wife or husband works “1” otherwise “0”
Job: Teacher: 0, Professor: 1, In the bank: 2, Engineer: 3,
Marketting: 4, doctor 5
If he or she has his own house “1” otherwise 0”
If he has a car “1” otherwise “0”
If he has a problem about his payments “1” otherwise “0”
FINAL DECISION: Positive “1” Negative (riscy) “0”
Machine learning:
The machine learning field evolved
from the broad field of artificial
intelligence, which aims to mimic
intelligent abilities to humans by
machines.
Learning types
Learning from the habits
Learning from the views
Learning from the commands
Learning from the samples
Learning by the way of analogy
Learning from the explanations
Learning from the experiments
Learning from the discoveries
In general, the machine learning systems are
designed for 2 purposes:
1- Hetereo-association: By using all learning
types mentioned above, trying to describe
the sistem generally based on learning. All
samples are evaluated and some
generalizations are obtained about the
problem especially in classification and
prediction problem
2- Auto-association: In this case one
event is learned and the data, the
system had, is used to characterise
the problem. In this case the system
can describe the missing points.
Human face recognition is an
example for this. The system can
also succeed if there is problem on
the image.
Learning Paradigmas:
Some paradigms and approaches are
improved to convert the studies
based on learning to machine
learning schemes. Some samples are
given below:
1- Symbolic procesing
2- Connectionist systems
3- Statistical pattern recognition
4- Genetic algorithms and evolutionary
programming
5- Case based learning
The Learning Strategies:
1. Supervised learning: A teacher is
needed here. The teacher should give to
the system input and output data for the
problem to be learned.i.e. Multilayer neural
networks.
2. Reinforcement learning: Here the
teacher just helps to the system. But,
instead of showing all outputs for the
inputs set, the teacher expects the system
to produce the output for the given inputs,
and produces a signal about the output is
true or false. Based on this signal, the
system goes on learning process. i.e. LVQ
networks
3- Unsupervised learning: There is
not teacher here helping to the learning of
the system. Here only the input values are
shown to the system, and the system is
wanted to learn the relationships among
the inputs by itself. This is mostly used for
classification problems. However, after the
learning process, the labelling has to be
done by the user. ie. ART networks are
using this strategy when it is learning.
4- Hybrit strategies: Any learning
strategy, which uses two or more of
these 3 learning strategies. Here, the
partially supervised or partially
unsupervised learning systems are
described. Radial bases neural
networks or probabilistic neural
networks can be given as examples.
Online and Offline Learning
On-line learning: The system can work
as a real time system. When the system is
working, the learning process can go on.
The rule in ART network and Kohonen
learning rule are from this learning
system.
Off-line learning: These systems are
trained off-line and then used. For any
new data the system is trained off-line.
i.e. Delta learning rule in neural networks.