Er is een explosie van toepassingen van Neural Nets en Deep learning. Wat kunnen deze wel en wat kunnen ze niet. Wat kan deze ontwikkeling voor U betekenen?
4. OUTLINE
1. Definitions
2. How does it work?
¡ Machine Learning Basics
¡ Deep Learning Basics
¡ Vision & CNN’s
¡ Practical tips to start with deep learning
3.AI @ Fontys ICT
4.Takeaways – What we learned from cases
6. 1. DEFINITIONS
¡ Artificial Intelligence
¡ Machine learning
¡ Deep learning
¡ Artificial Intelligence (AI) is the science of making
things smart. Can be defined as:
¡ “Human intelligence exhibited by machines”
¡ A broad term for getting computers to perform
human tasks. In practice, the scope of AI is disputed
and constantly changing over time …
7. 1. DEFINITIONS
¡ Artificial Intelligence
¡ Machine learning
¡ Deep learning
¡ Machine Learning (ML) can be defined as:
¡ “An approach to achieve AI through systems
that can learn from experience (data)”
¡ Machine learning involves a computer to recognize
patterns in data (i.e. by examples), rather than
programming it with specific (explicit) rules.
8. 1. DEFINITIONS
¡ Artificial Intelligence
¡ Machine learning
¡ Deep learning
¡ Deep learning (DL) can be defined as:
¡ “A biologically inspired technique for
implementing ML”
¡ DL code structures are arranged in layers (called
artificial neural networks) that loosy mimic the
human brain.Training these networks requires
massive amounts of data and computing power.
Object recognition
Speech recognition
Natural language
Processing
Translation
Signal restauration
(adversarials)
9. 1. DEFINITIONS – MACHINE LEARNINGVERSUS DEEP LEARNING
¡ Artificial Intelligence
¡ Machine learning
¡ Deep learning
Deep Learning
* More predictive power * Hard to train (many DoF) * Requires insane amount of data *
11. 2. MACHINE LEARNING BASICS – FEATURE ENGINEERING
How to distinguish
apples from oranges?
Idea’s?
¡ ...
¡ …
12. 2A. MACHINE LEARNING BASICS – FEATURE ENGINEERING
¡ Best practices for feature
engineering …
¡ Acquire domain knowledge
¡ Visualize
¡ Good features hold info and
are not correlated
¡ Consider transformations &
dimensionality reduction
(PCA)
¡ Note: data is fictive!
Bad features
colour
weight
#ofseeds
no yes
wrapped
Good features
14. 2. MACHINE LEARNING BASICS – FEATURE ENGINEERING
¡ Best practices for feature engineering …
¡ Consider measurement beyond our senses!
¡ Be creative, think out-of-the-box
Features from
expertsTNO
Each fish has its own
acoustic properties,
evident from the
amplitude and
structure of the
reflected energy.
Differentiate fish
species (like Salmon,
Herring, Mackerel,
etc) real-time.
15. 2. DEEP LEARNING BASICS
¡ Neural networks
¡ Perceptron
¡ MLP
¡ Neuron Network Zoo
¡ Vision & CNN’s
Neuron
Perceptron =
abstraction
of neuron =
decision unit
17. 2. DEEP LEARNING BASICS
¡ Neural networks
¡ Perceptron
¡ MLP
¡ Neural Network Zoo
¡ Vision & CNN’s
There are lot’s of NN configurations …
18. 2.VISION & CNN’S
¡ Vision is not easy !
¡ The eye (camera equivalent) is
only the starting point …
¡ The brain does heavy analysis
& processing
¡ This processing is needed to
create a coherent world that
makes sense and allows for
meaningful interactions with our
environment (e.g. navigation)
Animal eyes
19. 2.VISION & CNN’S
¡ Vision is not easy !
¡ The eye (camera equivalent) is
only the starting point …
¡ The brain does heavy analysis
& processing
¡ This processing is needed to
create a coherent world that
makes sense and allows for
meaningful interactions with our
environment (e.g. navigation)
Processing starts in the retina
Receptive field =
filter =
convolution kernel
21. 2.VISION & CNN’S
¡ Neural networks
¡ Vision & CNN’s
¡ How good are they?
Errorrateinimageclassification(%)
22. 2. PRACTICAL TIPS TO STARTWITH DEEP LEARNING
¡ Frameworks
¡ Pre-trained CNN’s
Don’t be a hero, use frameworks and
instantiate pre-trained nets !
23. 2. PRACTICAL TIPS TO STARTWITH DEEP LEARNING
¡ Frameworks
¡ Pre-trained CNN’s
Don’t be a hero, re-train the last layer
with your domain-specfic images !
24. 3.AI @ FONTYS ICT
¡ Fill-in declarations automatically
¡ Graduation project @ HRM Driessen
¡ Enhance (pre-process) receipts
¡ OCR with Tesseract
¡ NLP & machine learning to classify type
of declaration
¡ Integration in web-application
¡ Status: Finished July 2018
25. 3.AI @ FONTYS ICT
¡ For some people renting a
residence it might be
beneficial to buy one …
¡ Graduation project
@ DeVolksbank
¡ Find & mine the right features
in transaction data
¡ Privacy & security aspects
¡ Create working proof-of-
concept as demo
¡ Status: Finished July 2018
26. 3.AI @ FONTYS ICT
¡ When do cows become ill?
¡ Graduation project @ InfoSupport
¡ Find the right features
¡ Develop prediction model
¡ Create working proof-of-concept
as full-stack app
¡ Status: Started September 2018
27. 3.AI @ FONTYS ICT
¡ What is the quality of a room (house)?
¡ Cognizant case for ADS minor
¡ NVM image dataset
¡ Develop insightful model based on vision
¡ Make model assessable via API
¡ Status: Started September 2018
28. 3.AI @ FONTYS ICT
¡ When and where is (toxic) syntetic
drug waste dumped? Can you help us?
¡ Noord-Brabant case for ADS minor
¡ Police & province data is given
¡ Easy-to-use visualization (BI) tool that
provides insight in the existing crime data
¡ Prediction model (mapping with open data,
e.g. light pollution maps)
¡ Status: Finished February 2018
30. 4.TAKEAWAYS –WHATWE LEARNED …
¡ Concepting AI = Rethink your business
¡ Experiment to find added value= personalized
products, services, experiences
¡ Become data-driven
¡ Impact on processes, logistics, contracts, …
¡ Building AI = Like preparing a good meal
¡ You need many ingredients that fit together
¡ Build a diverse data science team
¡ Visualize & communicate results
¡ Consider legal and ethical aspects!