Introduction

AI Fundamentals

Artificial Intelligence

Terminator 2:
Judgment Day (1991)

Directed by James Cameron

2001: A Space Odyssey (1968)

Directed by Stanley Kubrick

Star Wars: Episode V - The Empire Strikes Back (1980)

Directed by George Lucas

Blade Runner (1982)

Directed by Ridley Scott

Her (2013)

Directed by Spike Jones

Ex Machina (2014)

Directed by Alex Garland

So, what is Artificial Intelligence (currently)?

The original vision

The original vision for AI was to build computer systems to mimic human intelligence to take over dangerous, dull and dirty tasks.

If we want to mimic human intelligence, AI needs to support:

• Processing
• Understanding
• Reasoning
• Taking action

AI building blocks

To go through cycles, an AI system requires three large building blocks:

• Algorithms
• Data
• Hardware

Instruction Based

• Programming instructions is pretty hard for a human
• Programming line by line, if situation A occurs, the system must take action B
• Takes a lot of time to write
• Very transparant

Instruction Based

Machine Learning

• Computers learn from data and adapt through experience, without a programmer defining the rules.
• Can respond to previously unknown situations
• Not transparant
• Most of what we consider as AI today is ML (Machine Learning)

But still...

Things can go wrong

Husky vs Wolf

Skin Cancer

I'm Fine

AI Systems

There are different ways to classify AI systems. The most common subdivision is:

Alan Turing

The Turing Test

[16:11:25] Judge: what music do you like?

[16:11:32] Eugene: I like Eminem – because he doesn't have a thing to do with "polite correctness".

[16:11:46] Judge: OMG i love eminem, he's my favourite artist!

[16:12:00] Eugene: I wonder what Eminem's nick mean. I associate it only with a word "enema". But he is a cool guy, I like his The Real Slim Shady.

[16:12:55] Judge: i know what his nickname means, his real name is Marshall Mathers. From the initials M&M phynetically sounding like Eminem.

[16:13:06] Eugene: Eminem is a famous crapper. No, raper… ripper… Yeah, I mean – rapper. I like his Stan and The Real Slim Shady.

Human-Centered AI

Humans are lazy by design, and that laziness is driving force for innovation. We need to be aware of:

• AI is a computer system
• AI is not a religion
• See AI as a third arm or a second brain
• Mor transparency from AI systems

Human in the loop

Overview

1. Algorithms
2. Data
3. Hardware
4. Automated Decisioning
5. Ethical AI
6. Intelligence

AI Fundamentals

Algorithms

AI Fundamentals

Data

Self Driving Car Emotions

So, what is an algorithm?

Rule-based algorithm

Self-learning algorithms

Sometimes, the number of rules becomes extremely long and complex, or is simply impossible to solve via rules

• Learning how to talk
• Recognising a cat in a photo

We use AI to mimic our brains, make automatic connections between input and output, drawing on examples

How?

Try to fill in the number where there is now a question mark

How?

We recognize a pattern and can state that
output = (input 1 x input 2) - input 3

How?

• Input 1, 2 and 3 are the attributes
  • More complex algorithms require more attributes
• A variable gives weights to an attribute

Compared to a recipes, the attributes are the ingredients, and the variables are the weights that determine how much of each ingredient we need.

Patterns

Neural Networks

• Neural networks are inspired by our brains and are a frequently used category of AI algorithms
• Inspired by neurons
• Complex adaptive system that can change the internal structure by data

Neural Networks

Deep Neural Networks

Training

Training

Training

0.1

0.9

0.9

Training

0.5

0.7

0.9

Neural Networks

Seems abstract but:

• YouTube has 30 layers to recommend new videos you might like
• Facebook uses one in DeepFace, which is 9 layers deep and processes more than 120 million input attributes and has been trained on four million images uploaded by Facebook users
  • 97% reliability
  • Auto tagging faces

Model

• When you train an algorithm on a collection of data, you get a so-called 'model'
• The same algorithm with different data, results in a different model

Metaphor: if we compare this with baking a cake, the recipe is the algorithm and the ingredients are the data. Different ingredients, different recipe... different model.

Model

Can our model create good predictions?

1. First we are in the training state, the system looks for the best possible connection pattern between the input and the output
   • In each iteration, the predications are compared with the desired ones, small adjustments are made to the weights
   • Training stops when the system indicates that the desired accuracy has been achieved

Model

2. We test the model to make predictions from data where we know the result, but the model does not. 

• If the model passes the test, its ready to be used. We call this the operational phase.
• If not, we need to review it

Model

• If our data is 97% accurate in training phase, but 50% with unseen data, we call this overfitting.
• If our data is 50% accurate in training phase, but 97% with unseen data, we call this underfitting. (Like as student guessing an answer).

Neuralink Monkey

Neuralink Tests

Learning

There are three major learning techniques for algorithms, which are used in combination with each other:

• Supervised learning
• Unsupervised learning
• Reinforcement learning

Supervised Learning

• Trained on a labeled dataset, the algorithm knows what the correct output is
• Tries to generalize to new examples that it's never seen before by
  • Classification, e.g. spam or not spam
  • Regression, the output is a continuous value, e.g. price
• More accurate than unsupervised, but a lot more work

Who labels data?

Unsupervised learning

There are no labels, the algorithm needs to discover hidden patterns in data without the need of human intervention

• Clustering groups similar experiences together. E.g. customer similliarities like a big spender or a window shopper
• Association looks for relationships between variables. E.g. a customer who bought this item, also bought...

No humans to intervene, they find patterns and can cluster data (no predictions!)

Unsupervised learning

Reinforcement Learning

We now the desired end goal of a problem, but we do not know how best to get there. Think of learning to ride a bike:

• We are rewarded when we keep cycling
• If we fall off we get punished
• The algorithm weighs each new action against previous knowledge

The system does not have examples, but learns by doing.

AlphaGo was trained using this technique.

Self Driving Car

Learn Robots How To Walk

Adversarial Learning

• Generative Adversarial Networks (GAN) is combination of supervised and reinforcement learning.
• There are two neural networks: 
  1. The discriminator: this is a network with a large dataset... let's say of cats
  2. The generator: this wil make its own output, for example creating pictures that look like cats (with the aim of fooling the discriminator)

Adversarial Learning

• First, the generator lies to the discriminator by random noise.
• Based on the prediction rate of the descriminator, the generator will try to create a better proposal
• The generator will create such good pictures that they are barley distinguishable from the originals

https://this-person-does-not-exist.com/

Reality check...

Self-Driving Car Crash

Hacking the road is very easy

Elaine Herzberg

Self-Driving Car Crash

Analysis showed that the Uber had seen Herzberg, but changed her classification several times between vehicle and cyclist. The system could not predict the path.

Let's be careful with criticism, there are a lot of humans who constantly make mistakes. We can use AI together to be safe, but still, self driving cars are not as narrow as AlphaGo.

Sensors

Using multiple sources like cameras, thermal camera, radars,... to get grip on our environment must collide data immediately in real time. This is called sensor fusion.

Algorithmic Bias

Weapons Of Math Destruction, Cathy O'Neil

#2 BIAS (2018) by playField.

Algorithmic Bias

Joy Buolamwini

Demo

Try and create image with leonardo.ai. Use prompts:

• Create an image of a grandmother
• Create an image of an engineer
• Create an image of a nurse
• Create an image of a boss/CEO

Value Added Model

Move 78

AI Fundamentals

Data

Introduction Data

Five Dimensions

Over the last ten years, we have seen an increase in five dimensions of data, all starting with the letter 'V':

• Volume or the amount of data
• Variety or the different types of data
• Velocity or the speed at which the data is available
• Value of the data
• Veracity or the quality of the data

Variety

Structured Data can be stored in a database or table. They are structured in columns and rows, similar to the way that spreadsheet software like Excel classifies data.

Variety

Unstructured Data cannot be stored in a traditional row-column database. E.g. photos, videos, sound files or large texts. They do not have a fixed data model.

Semi-structured data are somewhere between the two. E.g. photos that have metadata, information baked in the file like location, etc.

Velocity

1. The speed at which the system receives new data
2. The relationship between receipt and processing
   • Batch Processing
   • Near real-time-processing
   • Real-time processing

Value

• Data storage costs money
• Making decisions on what to store and try to keep the relevant data

Veracity

• High quality data are essential for high quality algorithms.
• Bad data gives biased AI systems

Tay AI

Consumer Data

We live in a highly connected world

• Smartphones
• Smart phone speakers
• Surveillance cameras
• Internet Of Things (IOT)

We need data to create a consumer profile for the best consumer experience

Personal

Data

Personal Device

Data

Online Data

GEO-location

Data

Social Media Data

Socio-demographic data & transactional data

Contact & Response History Data

The Customer's DNA

Recommandations

Amazon Go

#1 HOLLOW by playField (2016)

Social Credit System

GDPR

• The GDPR sets out the requirements for the management and security of personal data of European Citizens.
• Responsibility must come form two sides.
• You have the right to remove yourself out of data
• Accept Cookies

Solid