Some introductory words on

Artificial intelligence

Machine learning

Deep learning

Mattias Ohlsson

Department of Astronomy and Theoretical Physics, Lund University

School of Information Technology, Halmstad University
Artificial intelligence (AI), Machine learning (ML), Deep learning (DL) is
everywhere
In my own home...

“Natural Language Processing”

Everyday tools...

Google photo

Search term:
Katt
 Future …

Self-driving
cars

https://www.ericsson.com/en/blog/2020/1/self-driving-car-passenger-experience
 What is deep learning?

https://becominghuman.ai/deep-learning-made-easy-with-deep-cognition-403fbe445351

http://uc-r.github.io/feedforward_DNN

Google image search:

Deep learning

http://www.digitalrhetoriccollaborative.org/2018/04/23/inscrutable-ai-deep-learning-and-the-problem-of-technology-and-trust/
 https://www.zdnet.com/article/deep-learning-the-interest-is-more-than-latent/

https://codeburst.io/deep-learning-what-why-dd77d432f182

Google image search:

Deep learning

https://www.cgit.se/nyheter/vad-ar-deep-learning
 First, what about AI / ML / DL

Artificial Intelligence (AI)

Deep Learning (DL)

Machine Learning (ML)

 machines that could
successfully perform any
intellectual task that a
human being can

”the study of
intelligent agents”

Artificial General Intelligence (AGI)

Artificial Intelligence (AI)

Narrow AI

”the science and

engineering of
making intelligent
machines” recognizing cats
”my speaker” in images

self-driving cars
 Machine Learning

”Machine Learning is the study of

“Machine learning is the computer algorithms that
subfield of computer science learn and improve
that gives computers the ability automatically through experience.”
to learn without being explicitly
programmed.”
(Wikipedia) The field of machine learning is
concerned with the question of
how to construct computer
programs that automatically
improve with experience.
Models that learn (Tom Mitchell)
from data!
A simple machine learning task!
A more complicated machine learning task!
 A “famous” machine learning task!

Each digit is a 28x28 “grayscale”

image.

60 000 images in the training set.

10 000 image in the test set.

Best result so far are based on (deep) convolutional neural networks

or deep neural networks with an error rate of 0.21% (21 misclassified images)
 A lot of tools available for machine learning

Deep learning

Neural
networks

XGBoost Random forest

Eg. machine
learning for
classification Naive Bayes
Support vector
machines

Logistic KNN
regression
 We can consider ML a subset of AI

AI ML
 And deep learning a subset of ML

Deep Learning (DL)

AI ML DL
 To make it simple!

Artificial Intelligence Actual narrow AI

Machine learning is A lot narrow AI

today dominated solutions are based on
by deep learning machine learning

Deep learning is
actually a fancy
name for a collection
of tools that belongs to

Artificial neural networks

What can we do with machine learning?
 First, different kinds of tasks

Classification: The algorithm is asked to give one of

k classes for which the input belongs to

Separate red
from blue

Healthy or not?
 Regression: Predict a numerical output given an input

An example
●
Predict the useful remaining life of a battery
Translation: sequence in → sequence out

Time series prediction

Synthesis and sampling: : Generate new data “similar” to the given data

https://arxiv.org/pdf/1812.04948.pdf

Generate faces
Some application areas (growing by the day...)
Machine Classification
Translation
Object
Classification recognition
Natural
Language OCR
Images Processing

Generation
Speech Topic
(semantic) recognition segmentation
Segmentation Enhancement

Healthcare
Bioinformatics Data
security
High energy
physics
Financial Recommendations
trading
Smart
cars
Fraud Playing
Online detection games
search
Examples

Object detection
Semantic Image Segmentation Instance-aware segmentation
 Self-driving cars

Waymo has launched its robo taxi self-driving car service called
Waymo One in Pheonix, Arizona (USA) as of December 2018.

Video in youtube

Waymo has the best track record, recording over 10

million miles of testing without any serious accidents.
Different types of learning

Types of
Learning

Semi-supervised
Learning
●
Supervised Learning
Pattern mining given labeled data
●
Unsupervised Learning
Pattern mining in unlabeled data
●
Semi-supervised Learning
A mix of supervised and unsupervised
learning. There are labels for a subset
of the data, but not for all of it.
●
Reinforcement Learning
Learns to act based on
feedback/reward
●
Self-supervised Learning (new)
 Supervised Learning
 Supervised Learning involves using labelled data sets that have inputs and expected outputs, i.e.
ground truth.
 When you train a machine learning model e.g. a CNN, using supervised learning, you give it an
input and tell it the expected output.
 If the output generated by the model is wrong, it will readjust its calculations. This process is done
iteratively over the data set, until the model makes no more mistakes.
 Example: Image classification with CNNs

CNN
 Unsupervised Learning

 Unsupervised Learning uses data sets with no specified label.

 The machine learning models will look for patterns in the data.
 Example: Clustering or Autoencoders
 Unsupervised Learning

“Compressed representation”

Encoder Decoder

Autoencoder
 Semi-supervised Learning

 Semi-supervised Learning is a mix of supervised and unsupervised learning

 Uses typically a small amount of labeled data with a large amount of unlabeled data.

An example
 Semi-supervised Learning

Use unlabeled data to obtain

efficient data representations
(plenty of unlabeled data)

Use these representations when training

the machine learning model (few labeled data)
 Reinforcement Learning
 Reinforcement Learning is a method that enables an agent to learn in an interactive
environment by trial and error using feedback (reward) from its own actions and experiences.
 An agent uses rewards and punishment as signals for positive and negative behavior.
 The goal is to find a suitable action model that would maximize the total cumulative reward of
the agent.
 No explicit training data set is needed, i.e. more general than supervised/unsupervised learning
 Today machine learning methods based on
neural networks are very popular!

Multi-layer Convolutional
perceptron neural networks

Long short- Neural Boltzmann

term memory Networks machine

Generative
adversarial Autoencoder
network
 A simple neural network (MLP)

Input Output

A mapping from “input” to “output”

Tunable weights
Many different kinds of network architectures
http://www.asimovinstitute.org/neural-network-zoo/
Some common architectures

Multi Layer Perceptron (MLPs)

http://uc-r.github.io/feedforward_DNN
Some common architectures

Convolutional Neural Networks (CNNs)

Some common architectures

Deconvolutional Neural Networks

Some common architectures

Autoencoder Networks
Some common architectures

Recurrent Neural Networks (RNNs)

Some common architectures

Generative Adversarial Networks (GANs)

Some words about deep learning

ML DL
 Deep learning

Representation Layer by layer

learning learning
 Circle or rectangle?

What property is important for this question?

 Largest circumference?

What property is important for this question?

 Layer 1 Layer 2 Layer 3

that is used to build that is used to build even

Simple features
more complex features more complex features

Each
Each layer
layer is
is learning
learning efficient
efficient features
features for
for the
the question
question at
at hand.
hand.