SkillsBuild

Machine Learning
and Deep Learning Essentials

Images used in this presentation by the trainer stem from the Canva Pro license acquired by the trainer.
Brief introduction for
newcomers
ReCap of last live session
Machine Learning Essentials
Exercise 1
Deep Learning Essentials
Career in this realm
 Why am I teaching you?
Some very quick background on myself...
Kristina Bodrožić-Brnić

PhD project on the connection of GenAI

with technology openness in SMEs
Current functions:
AI Trainer at SkillUp
Project management for Creative AI and AI Trainer at the Mittelstand-
Digital initiative of the Federal Ministry for Economic Affairs and
Climate Action of Germany
Project management & teaching at the IDTI of the BSP Business and
Law School
Head of Operations at WorldTribe
How did I start with this topic?
Types of interaction:

Q&A Sessions (ask questions you might have and I

ask mine)
Experimenting alone and giving feedback (if with
other in front of the screen, involve each other)
My first question:

From which continent are you doing this course?

(1) Asia
(2) Australia
(3) Africa
(4) Europe
(5) South-America
(6) North-America
ReCap of last live session
AI has different names and history when it comes to development
People often feel fear or overestimation when confronted with AI tool usage
There is weak AI (what we use now) and strong AI (independent human-like decision-making) only
exists in Science Fiction
We call technology everything mankind has developed to create something that aids them to
perform a task better (stone tools)
Human-centred AI development want just that, that tech continues serving us
This is why we do this course. We want to know some basics on AI. We increase our knowledge
about AI. We all increase “AI Literacy” in our societies
Two fathers of AI were Alan Turing and Geoffrey Hinton
NLP stands for Natural Language Processing (providing computers the ability to process data
encoded in natural language = recognition and understanding, and output of natural language),
uses primarily artificial neural networks since 2015)
GenAI means generative AI and uses NLP. We tried to “interact” through “prompts” last time and
experimented with text-generation and text-to-image-generation

Trainer: Kristina Bodrožić-Brnić

 Machine Learning (ML) Esstential

Classic programming
Three basic types of Machine
Learning
Supervised Learning
Unsupervised Learning
Reinforced Learning

Trainer: Kristina Bodrožić-Brnić

Machine Learning (ML) Esstential

Artificial Intelligenc e(AI)

Image source: Mittelstand-Digital Zentrum Kaiserslautern

Trainer: Kristina Bodrožić-Brnić

 Classic Programming
results are
Often not
explainable
labled AI

Rules classic
Answers
Data programming
The implementation of AI-based solutions is based on algorithms. An algorithm is a series of
instructions executed step by step to complete a specific task.
In classic, rule-based programming, the developer uses algorithms to specify rules to the
system to which the input data is applied.

Trainer: Kristina Bodrožić-Brnić

 Machine Learning (ML) Esstential

Self-learning AI AI
More flexible: sometimes one algorithm
can be used for different problems
sometimes labeled as AI, sometimes not,
“automation”
often not clearly explainable, can show
unexpected behavior

Trainer: Kristina Bodrožić-Brnić

 Supervised Machine Learning

Labeled Data: The training dataset includes both inputs and their corresponding correct
outputs (labels).
Guidance: The algorithm uses these labels to guide the learning process, adjusting its model to
minimize the difference between its predictions and the actual labels.
Evaluation: The performance of the algorithm can be evaluated using the labeled data,
providing a clear measure of how well it is learning.

Supervised ML models were often criticized in image recognition because

such models usually perform worse than 5-year-olds. However, in other
kinds of automation processes, this can be very useful.
cat cat cat ? ? ?

Trainer: Kristina Bodrožić-Brnić

 Supervised Machine Learning
Linear regression: This is a statistical technique used to determine the relationship between a
dependent variable (often called the target) and one or more independent variables
(predictors). Essentially, it aims to fit a straight line through the data points to make predictions
or estimations. For example, you can use linear regression to estimate house prices based on the
size of the house in square meters.

“In linear regression, the observations (red) are

real value

assumed to be the result of random deviations (green)

from an underlying relationship (blue) between a
dependent variable (y) and an independent variable
(x)”.

square metres

Image Source: https://en.wikipedia.org/wiki/Linear_regression

Trainer: Kristina Bodrožić-Brnić

 Supervised Machine Learning
Logistic regression: This supervised machine learning algorithm is used for binary classification
tasks. It predicts the probability of a specific outcome or event, such as whether it will rain today
(yes or no). The result is a value between 0 and 1, representing the likelihood of the outcome.

Image Source: https://en.wikipedia.org/wiki/Linear_regression

Trainer: Kristina Bodrožić-Brnić

 Supervised Machine Learning
Decision Tree: This is a predictive model used for classification and regression tasks. It works by splitting a
dataset into branches to draw conclusions from a set of observations. Commonly used in data mining, the
goal is to predict the value of a target variable based on multiple input variables. For example, it can be
used to determine the conditions under which tennis was played and predict based on this.

Trainer: Kristina Bodrožić-Brnić

 Term: Data Mining
Data mining involves exploring large datasets to uncover patterns and relationships that can provide
insights and help address business challenges through detailed data analysis.

Trainer: Kristina Bodrožić-Brnić

 Supervised Machine Learning
The k-nearest neighbors (k-NN) algorithm is a non-parametric, supervised learning method that
classifies or predicts the category of a data point based on its proximity to nearby data points.
Lazy Learner = all trainings must be saved; doesn’t memorize; no training phase; expensive
because learns all over through the whole set.

Example of k-NN classification: Consider a

test sample represented by a green dot
that needs to be classified as either a blue
square or a red triangle. If k = 3 (solid line
circle), the green dot is classified as a red
triangle because there are 2 triangles and 1
square within the circle. However, if k = 5
(dashed line circle), it is classified as a blue
square because there are 3 squares and 2
triangles within the larger circle.
Image source: https://en.wikipedia.org/wiki/K-nearest_neighbors_algorithm
 My question:

Why do we call supervised learning “supervised”?

 Unsupervised Machine Learning

Clustering:
“In machine learning too, we often
group examples as a first step to
understand a subject (data set) in a
machine learning system. Grouping
unlabeled examples is called
clustering.” (market segmentation,
medical data, anomaly detection)
finding interesting groups (are their
unknown aspects my clients have in
common, so we can improve
marketing?)
results from same and different
algorithms might find new results Image source: https://www.sciencedirect.com/topics/computer-science/mean-shift-clustering
 Unsupervised Machine Learning

Clustering:
Similarity measure, like books from
specific authors the customers read.
clustering data can simplify large
datasets. As said, such data is unlabled.

Image source: https://developers.google.com/machine-

learning/clustering/overview#:~:text=In%20machine%20learning%20too%2C%20we,relies%20o
n%20unsupervised%20machine%20learning.
 My question:

Where do you encounter recommendation-systems?

 Unsupervised Machine Learning

Recommendation systems:
 Unsupervised Machine Learning

Collaborative Filtering

Needs large data-sets

Needs users to “share” data (collaborative)
“Typically, the workflow of a collaborative filtering system is:
A user expresses his or her preferences by rating items
(e.g. books) of the system. These ratings can be viewed as
an approximate representation of the user's interest in the
corresponding domain.
The system matches this user's ratings against other users'
and finds the people with most "similar" tastes.
With similar users, the system recommends items that the
similar users have rated highly but not yet being rated by
this user”
Image Source:
https://en.wikipedia.org/wiki/Collaborative_filtering#:~:text=In%20the%20mo
re%20general%20sense,involve%20very%20large%20data%20sets.
Unsupervised Machine Learning
 Reinforced Machine Learning
also called “neuro-dynamic programming”

“Reinforcement learning (RL) is a

machine learning (ML) technique
that trains software to make
decisions to achieve the most
optimal results. It mimics the
trial-and-error learning process
that humans use to achieve their
goals.” A bit like with children! It
tries something out and gets
feedback. By the way it received
feedback (immediate, slow,
negative) it learns to connect the
dots.
 Deep Learning
complex pattern recognition, therefore it likes “complex sencsoric data” such as audio,
language, images, and needs very large data
it mimics human way of learning and output
compared to linear regression, it finds more connections, creates more nodes (a point in a
network or diagram at which lines or pathways intersect or branch.)
Linear regression
DL finds more connections
such as location, income,
number of family members,
quality of schools in the
neighborhood.
 Deep Learning
Input layer, one or more hidden layers, and an output layer.
transforms the information to the next layer
 Deep Learning
Uses multiple layers, they form neural networks
applied e.g. in computer vision and speech recognition
examples of car bumps simulation, which areas are being activated?
 Deep Learning
Convolutional neural networks (ConvNets or CNNs) are more often utilized for classification
and computer vision tasks, for example as follows
the hidden layers are convolutional layers, but can contain in addition normal layers
uses filters, that detect specific things (eyes, feathers, dogs, cats...)
divides the picture into squares and the filter goes over it (convolves) and has an output that
resmbels the original and send it to the next layer and repreats
segmentation object recognition

https://images.app.goo.gl/xPTzZSL6ZuXJ7tYQ9
 Deep Learning
Very simple example of convolving filters in CNNs!

filters detect edges

different filters such as circle detecting filter

Source https://www.youtube.com/watch?v=YRhxdVk_sIs
Deep Learning - Recurrent Neural Networks (RNNs)
Use in forecasting, e.g. stock-orders or word by word generation.
Gos in two ways.Output of individual layers can be fed back into the previous parts.

To predict where the ball is going we need more than just access to the current position of the ball.

The sequence we get tells us the ball is moving right.

Text is also a sequence. RNNs have a consequentional memory. “ I love ___.” (you, dogs, burgers, ...)

Loop mechanism between Short term memory.

the steps within
the hidden layer helps =
hidden state from
previous inputs.

Source https://www.youtube.com/watch?v=LHXXI4-IEns
Deep Learning - Generative Adversial Networks (GANs)

Portrait of Edmond De Belamy, an AI-generated portrait by the

French collective Obvious
AI-generated art for $432,500

https://aws.amazon.com/de/what-is/gan/

The generator neural network identifies data attributes from the training set, while the
discriminator analyzes the same data to distinguish between attributes. The generator then
adds noise to some attributes and passes this modified data to the discriminator, which
calculates the probability that the generated data is real. The discriminator provides feedback
Source https://qz.com/quartzy/1437876/ai-generated- to the generator to reduce noise in future iterations. The generator aims to maximize the
portrait-of-edmond-de-belamy-sold-for-432500
discriminator's error, and the discriminator aims to minimize it. Both networks evolve through
training iterations until the discriminator can no longer distinguish between real and
synthesized data, marking the completion of the training process.
 GAN Use-Case

Image Super-Resolution:

Problem: Low-resolution images lack detail and clarity, which

can be a significant issue in applications like satellite imagery,
medical imaging, or old photographs.
Use Case: GANs can be used to enhance the resolution of these
images. For instance, in medical imaging, GANs can convert
low-resolution MRI scans into high-resolution images, helping
doctors to make more accurate diagnoses.
 GAN Use-Case

Synthetic Data Generation for Training Machine Learning Models:

Problem: Training machine learning models often requires a large
amount of labeled data, which can be expensive or difficult to
obtain.
Use Case: GANs can generate realistic synthetic data that can be
used to augment training datasets. For example, in autonomous
driving, GANs can create diverse traffic scenarios and
environments to train self-driving car algorithms more effectively.
 GAN Use-Case

Style Transfer in Art and Design:

Problem: Some visual artists and designers often seek to apply
the style of one image (like a famous painting) to another image
(like a photograph), but doing this manually is time-consuming
and requires significant skill.
Use Case: GANs can automate style transfer, allowing users to
apply the artistic style of a famous painting to a photograph with
high precision. This can be useful in digital art creation,
advertising, and even in creating themed visual content for
movies and games.
 GAN Use-Case

Facial Recognition and Image Editing:

Problem: Modifying specific attributes in images, such as
changing facial expressions, age, or gender, is complex and
challenging to do manually.
Use Case: GANs can edit facial attributes in photographs. For
instance, the FaceApp application uses GANs to alter facial
expressions, make a person look older or younger, or even change
their gender. This technology is also useful in the entertainment
industry for visual effects.
 GPTs
Generative Pre-trained Transformer
Python is a high-level, interpreted
programming language known for its
readability, simplicity.
Python is the cornerstone language for
developing, training, and fine-tuning GPT
models due to its extensive ecosystem of
machine learning libraries and frameworks. For
deployment and integration into applications,
a mix of languages including and various web
technologies are used to create a complete,
functional system.
 GPTs

Own intuitive GPTs, just a quick demonstration.

INTUITIVE EXERCISE

MY EXPERIENCE
INTUITIVE EXERCISE
 CAREER IN AI

engineering (AI, ML, data, robotics, software)

AI project lead in an SME (with some tech
background)
AI expert in HR integrations in an SME
AI program driver in any kind of business or
public sector positions (basic know how and a lot
of interest)
[email protected]