2. Machine Learning
ML is a method of data analysis that incorporates the utilization of algorithms
that have the capabilities to learn from data without programming
Computer science svvvv AI (a branch CS) ML (application of AI)
3. Machine Learning: GET SIMPLE!
The world is full of data. Then use algorithms to predict and work for you. YOU
CAN NOT AVOID ML
Machines and computers now do not need to have every detail programmed.
They can be programmed to learn by themselves and to work more efficiently.
Human mind is multitasking and flexible. Computes are not yet there but each
year they are capable of doing more and more complex things
4. Neural Networks
Neural Networks can looks like science fiction but is too real.
NN are a series of mathematical formulas (algorithms) mimicking human brains
NN can constantly adapt to new information and changes
Using NN computers constantly adapt and don’t follow the always the same
instruction (mimicking brains- At least some brains!!)
5. Neural Networks -Brain
HUMAN GAIN
EXPERIENCE BY
THE 5 SENSES
THE BRAIN PROCESSES THE
INFORMATION AND
COMPARE WITH PREVIOUS
EXPERIENCES
CONLCUSION
SIMILAR TO
ML AND NN
PASSWAYS
OF THINKING
are constantly
shifting
between
billions of
neurons
6. DATA, KNOWLEDGE, ALGORITHMS
Data is one resource that continues to grow
AI turns Data into knowledge since in the past human being would be needed to
analyze large amounts of data.
Algorithms – instead of a rigid code the data becomes parts of an algorithm and
the machine creates its own reasoning
Example of algorithm: the email can be send to spam, social etc based on
different patterns that are constantly chancing
7. 3 Types of ML
SUPERVISED LEARNING – requires an input of data. Supervised because it requires human input to ensure the computer has the
right data. The goal is to learn from “labelled data”.
(Classification - goal of predicting under which class new instances would be place e.g. spam email)
(Regression – predict of continuous outcomes e.g. Analyzis of the length of study for SAT and the score for a population of students)
UNSUPERVISED LEARNING – the computer does not have all the information to make a decision. May have some data.
(Clustering – technique which let us organize a great amount of data into subgroup or clusters. E.g. marketers use this to determine
specific customer groups)
(Dimensionality Reduction – compress the data sot it fits into a smaller subspace while keeping the relevant information intact)
(Semi-supervised –blend of supervised and reinforcement learning).
REINFORCEMENT LEARNING – goal is to develop a system that can improve its performance. The computer knows when to move
but will readjust it when it does some mistakes. E.g. Chess (the move depends on the environment).
8. PROCESS/training models
Preparation (pre-processing) the raw data is the most important step of any ML
application.
Important to choose the correct model between different algorithms: How can we
do that? We should have a training dataset and a validation dataset.
The parameters should only be obtained from the training dataset and then
applied to the test dataset.
10. ISSUES/PROBLEMS OF ML
ML is still evolving and still being used every day
E.g.1 Natural Language Processing – because different languages and accents –
challenges for Cortana, Alexa and Siri.
With stabilized reinforcement learning the ability to control robotics would open
up endless possibilities
GAN-Generative Adversarial Networks-set of algorithms implemented in 2 NN
working against each other. But they can crash!
ML does not have many problems but it is still in its infancy
11. Problems with ML not yet solved
Variable event space
Processing languages – there is no context understanding yet.
Facial identification –not yet perfect e.g. doesn’t recognize a female with makeup
Automated learning is a big part of ML: Leaning has to come from a variety of
resources. E.g. IBM Watson
3 way human rule not yet achieved by ML: take an image, gives it a type and then
gives it a description.
A lot of memory space is need to collect all the data
12. Examples of problems solved
Most of email spam actually goes to spam
ML can learn about your credit card spending and if something is not usual can
alert as stolen
ML can read post codes in envelopes
Facial recognition to unlock the device, passports at the airport
Similar suggestions by Amazon (Website algorithm)
Babylon for symptoms
Stocks: when to hold, buy or sell
13. Examples of problems solved
Most of email spam actually goes to spam
ML can learn about your credit card spending and if something is not usual can
alert as stolen
ML can read post codes in envelopes
Facial recognition to unlock the device, passports at the airport
Similar suggestions by Amazon (Website algorithm)
Babylon for symptoms
Stocks: when to hold, buy or sell
14. Where ML is used
Autopilot airplanes
Search engines
Language translation
Facebook newsfeed is personalized for you as well as You Tube, Amazon
IMPORTANT: If you have a large dataset in your organization, high quality and
then use it to improve.
15. The concept of Neural Networks
Similar to neurons in Brain, with inter-conextions.
Several layers of this neurons and each layer has its own purpose.
TYPES OF NN
A) CONVOLUTIONAL NEURAL NETWORKS (CNN)
Your brain can instantly recognize a photo of your mother or a bird for instances. For computers
this is complex and they have to use CNN
Used for image recognition, because it adapts the algorithms.
B) PERCEPTION (binary function that produces 1 or zero) AND BACKPROPAGATION ( gradient
moves from the last layer till the first layer. Basically goes back to correct the errors. It is a trial
and error method. Uses successive iteration (or attempt).
16. Neural Networks and AI
AI is the science of developing machines that have the ability to constantly
increase their intelligence.
Computers can learn but not think. To try to overcome those obstacles a unique
form of ML has been developed – this is AI
Some tasks such as identify text is so commonplace that is no longer consider AI.
Computers said to have AI are those that are able to play chess, self-driving-cars
etc
17. AI and Medicine
E.g. IBM Watson -cognitive computer used for healthcare.
Areas:
- Data management
– choose the right therapy for cancer
- -The Deep Genomics – doctors can find what is inside the cells
- -speed up the clinical trials
- -recognizing patterns of radiology images
18. AI and Finance
E.g. Weathfront – analyses the patter of spending and gives advice
JP Morgan uses AI to extract important data from legal documents saving
annually 360 000 hours of manpower
AI is in most of the finance companies to save money and increase revenue
AI used to calculate parameters beyond human capacities
Get important information from contracts
Inspections that goes beyond human accuracy
- COUNTLESS APLICATIONS
19. AI in translation and game playing
TRANSLATION
Machine makes clusters of words e.g. trees and leaves and this pattern is similar
in different languages.
A way to test and learn is translate from language A to B and than back from B to
A. If it is different the machine self corrects.
GAMING
The gaming world is the safest environment where scientific problem solving can
work in harmony with creativity.
Using AI two players can be in different part of the world.
20. Deep Learning
Sub-type of AI using Neural Networks
With Deep Learning the computer is capable of making continuous adjustments
in order to improve its performance every time it is used
E.g. Google to predict what websites you will must likely to visit
Deep learning – e.g. automatically create a record of the number of vehicles that
pass thought a certain point on a public road (categorize the model, the maker,
the colour etc –difficult for an human to be standing taking care of the task.
Pattern recognition – e.g. to expand the internet of things by collecting data from
any device that is connected to the internet.
21. Challenges
The most advanced NN are pretty primitive when compared to the human brain
Delta Rule- most common rule to train a NN. The input and output vectors are
compared and results adjusted using different algorithms that are already in the
system.
22. IMPORTANT NOTE
Machine Learning for Beginners. Guide to understanding Machine Learning.
Matthew Kinsey. 2018.
This presentation is a brief summary using sentences of the book above and is
used without any permition of the trademarks, brands or authors. This is for
clarifying purposes only.
Do you want to join a Deep Learning Research Project/Start-up in Radiology:
email me please: afernandes@s24global.com