Deep Learning Applications

-1-
Outline
• Introduction into AI
• Fundamentals of machine learning
• Introduction into deep learning
Introduction into AI
• General Definition
– Field of computer science that aims at creating machines
able to simulate human intelligence, it relies on algorithms to
analyze large amounts of data, identify patterns, and
perform tasks that would normally require human cognitive
abilities
• Human intelligence indications: reasoning, learning, problem solving,
decision making, lge understanding
– Verbal comprehension: similarities, vocabulary, information
– Perceptual reasoning: block design, matrix reasoning, visual puzzles
– Working memory: digits, arithmetic, …
– Processing speed: symbol search, coding
• Human intelligence definition: Mental quality that consists of the
abilities to learn from experience, adapt to new situations,
understand and handle abstract concepts, solve problems, and use
knowledge to adapt and manipulate one’s environment.
Introduction into AI
• Types (1): weak/ strong/ superintelligence
– Purely reactive:
• Do not have any past memory or data to work with, Do not form
memories neither , Take actions/ react based on what they see (told
to), Systems only react, Do not use any past experience for new
decisions
• Specialized on one field of work only: washing machines
– Limited memory
• Systems look into the past, Use previous data and keep adding it to
their memories over period of time, It has enough memory or
experience to make proper decisions, Information is short lived,
Neural networks for an object X identification, do not remember all
the pictures being entered while traning, only remember the
patterns extracted defining the object X

– Theory of mind
– self-aware
Introduction into AI
• Types (2): weak/ strong/ superintelligence
– Theory of mind
• Have the capacity to understand thoughts and human
emotions and how they affect decision making
• Adjust their behaviour according to human understanding,
Interact socially successfully!?

– Self aware
• The future generations of machines: Systems being aware of
themselves, understand their own internal states, Predicting
other people s feeling and act appropriately
• Eg: having a companion who would know whether the human
is happy (understand happiness and effects) and suggests a
cup of coffee to cheer him/ her up based on his/ her behaviour
analysis and multidimensional circumstances
Introduction into AI
• Applications (1)
• Personalization and recommender systems: work within the same logic of recommending relevant
items to the costumers based on their preferences (quite broad range of app)
• Eg1: E-commerce: Amazon: recommended products based on assignment needs + history
• Eg2: Streaming services: Netflix: preferences are analysed and suggestions are made of what one may like to watch or listen is
• Eg3: Travel booking systems: Booking: analysis of search history and behaviour to offer matching accommodations and flights
• Speech recognition and generation: recognize human speech and decide an answer: 1speech
recognition converted to a text format, 2text is analysed to get an appropriate output, 3 speech
generation algorithm reads the answer
• Eg1: Google Assistant (+ Duplex): artificial assistant which proposes several routines and makes tasks as calling and searching
and reminding … through applications on the phone, PC, home embedded S
• Natural language processing: understanding the natural language by 1extracting entities (names,
objects) and defining intentions (Calling/ ordering actions), then 2fulfil the request
• Eg1: Opinion mining: Google translator, Albert AI: find positive and negative sentiments in texts
• Eg2: Sentiment analysis: Rankers systems analysing brands or products texts/images in social media
• Computer vision: able to see in the AI field is to train machines to analyse images and videos and
recognize objects on them
• Eg1: Instagram filters: recognising human faces and cover them with the filter to be applied
• Eg2:: Identification: mobiles lock , surveillance cameras detecting criminals and crimes, … Google Lens
• Eg3: Health care: Recognizing tumours or risky conditions in MRI/ Xray images
• Eg4: Self driving: using images and videos recognition (cars, bicycles, pedestrians, road marks), lighters (distances), radars,
Introduction into AI
• Applications (2)
• Revenue optimization: many metrics used by ML in marketing and sales:
• Advanced sales forecasting: using market data and some demographic trends to train algorithms to make precise forecasting
• Churn rate: to know whether a customer will leave and when by analysing the subtle changes in the customer behaviour
• Customer lifetime value analysis: based a handful of user actions
• Dynamic pricing: airline, travel industries: come up with the self best prices to maximise the revenue by analysing current
market and demand taken into account historical data and new trends (these prices change dynamically multiple times a day
reacting to market conditions)
• Price prediction:
• Engines that are used in customer based application in order to help customers defining the appropriate when moment to
make a purchase
• Internal analytic systems for highly volatile commodity prices to make decision on buying goods (like electricity, oil, wood) or
even setting up prices based on competitors
• Fraud detection:
• Banks and other payment organizations (and for insurance to find fake claims) by basically looking for suspicious transactions:
comparison of customer incomes and what this latter spends periodically → blockage
• Ecommerce: FDA help to detect fake products reviews and reveals common urgencies
• Predictive maintenance: predicting failures in heavy machinery: when operating turbines factory
boats or some heavy machines and gather data from their sensors, predictions when machines will
fail based on some early signs is possible → The maintenance schedule should be adjusted and
repairs are forced in advance. ( transportation, robotics, agriculture, …)
Introduction into AI
• Components
– Advanced algorithms: brains which process and execute
– Large amounts of data to learn and function
– Computational infrastructure: hardware and software
framework required to support the execution of
algorithms
Introduction into AI
• Key features and objectives of AI
– Imitation of human intelligence
– Machine learning and adaptation
– Complex date analysis
– Autonomous decision making
Introduction into AI Let’s imagine
how ai changes

• Challenging! the world !

Fundamentals of Machine Learning
• Definition
• Traditional programming:
• Implementing an algorithm (a consecutive completed sequence of instructions allowing the resolution of a particular problem):
expression of a systemic method for the resolution of a determined problem, must be interpretable by a machine.
• Dictates strict and explicit rules to machine: Dijkstra algorithm
• Rules based systems: never do things they are not told to do
• Problem!: explicit programming does not work in many cases!
• Considering the Eg
• People are good at recognizing objects in real world and images: horse/ monkey
• Impossible to build a machine (system) which does that as perfectly as humans traditionally
• Imagine all the rules to be written to recognize a horse in a picture
• Elaborating all horse features: different horses, colours, sizes, angles, contexts, dimensions
• Recognizing voices, words by people easily!

• Machine learning allows algorithms to solve such tasks as well

• Does not use traditional explicit programming
• Instead: machines are built to come up with rules and ways to solve problem themselves

Machine learning is the technical basis of finding patterns in data. It

includes many methods and mathematical models for approximating
real-world problems from which the data is collected
Science of making computers learn and act like humans by feeding data and
information without being explicitly programmed: learn, predict, improve
Fundamentals of Machine Learning
• Definition
• ML allows taking lots of data and let the machines recognise the dependences and patterns in those
data in order to solve a particular problem
• The process is so called training
• ML can be understood by 2 major components:
• Use algorithms to find meaning in random raw and unordered data
• Use learning algorithms to find relationships between that knowledge and improve that learning process
• ML do that by learning from data, and producing models been trained to solve problems

• How machines learn about a given subject:

Training input Learning algorithm

Input (personalised) model output

->We have to put some training input and put that into a learning algorithm
Which will give us some knowledge
-> That knowledge will be what the machine knows about that specific subjects
Fundamentals of Machine Learning
Google translate
Google learning artificial neural network algorithm
Training input Learning algorithm
Google data
Input (personalised) model output
"Can you show me the way Google translate model "Pouvez-vous
to Hilton Hotel, please?" me montrer le
chemin de
l'hôtel Hilton,
• Process s'il vous plaît ?"

Google image recognition (pure arab blood horse)

Horses DB Clustering learning algorithm

Training input Learning algorithm

Input (personalised) model output

Horse classifier model Class features
Fundamentals of Machine Learning
• Types (1)
• Supervised learning: a method used to enable machines to classify/ predict objects, problems or
situations based on labelled data

• Take a set of data and add labels to it with the right answers to the problem
• Gather thousands of images of cats and dogs, manually label them, then feed the labelled data to a learning
algorithm
• This algorithm learns about the specifities of cats and dogs in this data by itself (the more data are used the
better the machine learns the results)
• The answer is already known

o Systems able to predict future outcomes based on the past data

o Require both an input and output to be given to the model for it to be trained

• Eg: identify a bicycle in a picture holding also a car and a bus

Fundamentals of Machine Learning
• Types (1)
• Supervised learning:
Fundamentals of Machine Learning
• Types (2)
• Unsupervised learning: Systems able to identify hidden patterns from the input data provided

• Data are not helped with labels

• Instead, machines are suggested to process data by themselves
• Group items by similarities
• Find anomalies
• It will not know that there are cats and dogs but it will provide pictures into groups withing features of interest

o By making data more readable and organized, the patterns, similarities and anomalies become more
evident

• Eg: separate this picture for things that do not match!

Fundamentals of Machine Learning
• Types (2)
• Unsupervised learning:
Fundamentals of Machine Learning
• Types (3)
• Reinforcement learning: important type of ML where an agent learns how to behave in an environment
by performing actions and seeing results
• used in very specifically: when algorithms have to deal with environment: agent (environment + action) -> new
state
• Robots and gamers profiles are good examples
• To train such
• The AI starts randomly doing all different actions
• If it does smth right by gaining points or jumping through obstacles -> it receives a reward
• If it fails -> it dies or gets punished
• After a long trial / error process -> the AI learns how to effectively interact with its envirnmenet
• World champion of GO was defeated this way by alphaGo AI (the google deepmind project, a computer program
that played the bord game GO ) (May 2017)
• Using games to learn how to learn and not another aspect

o Systems are given no training

o They learn on the basis of reward/ punishment that was received from the performance of their last
action
o Help increasing the efficiency of a tool, function, or program