Basic Computer Skills

(Artificial Intelligence and Machine Learning)

Week 3
Lecture 5 - Machine Learning, Deep Learning,
Ethical Issues and Other Problems.
 Outline
Introduction to Machine learning...............................................................................................................3
A Story by Google Bard.......................................................................................................................... 3
What is Machine Learning?.....................................................................................................................4
Definition of ML............................................................................................................................... 5
Types of Machine Learning.............................................................................................................. 5
● Supervised learning:................................................................................................................5
● Unsupervised learning:............................................................................................................6
● Self-supervised learning:.........................................................................................................7
● Reinforcement learning:..........................................................................................................9
Applications of Machine Learning................................................................................................... 9
Discussion Points and Questions.................................................................................................... 10
Machine Learning Models [Optional]................................................................................................... 13
K-nearest neighbors (KNN)............................................................................................................ 13
Decision Trees.................................................................................................................................14
Artificial Neural Networks..............................................................................................................15
Deep Learning.................................................................................................................................18
Generative Models.......................................................................................................................... 19
Benefits of Machine Learning............................................................................................................... 20
Distinguishing the terms ML, DL & AI................................................................................................ 21
Why is AI a big deal?.......................................................................................................................... 22
Why now?..............................................................................................................................................22
Ethical issues and other Problems with AI.............................................................................................. 24
Discussion Points...................................................................................................................................27
How to build AI systems?.......................................................................................................................... 28

2
Introduction to Machine learning
A Story by Google Bard
Imagine you're a parent teaching your child how to identify different types of
animals. You show them pictures of various animals, pointing out the
characteristics of each one, such as a dog's floppy ears, a cat's sharp claws, or a
bird's wings. As you show them more examples, your child starts to recognize the
patterns and can eventually identify new animals on their own.

This is similar to how machine learning works. Instead of a parent teaching a child,
we have a computer algorithm learning from data. We feed the algorithm a lot of
examples, labeled with the correct answer, and it starts to recognize the patterns
that separate different types of data. Once the algorithm has learned these patterns,
it can be used to make predictions about new data that it hasn't seen before.

There are two main types of machine learning: supervised and unsupervised. In
supervised learning, we provide the algorithm with labeled examples, meaning the
data is already tagged with the correct answer. This is like teaching a child by
pointing out the different characteristics of each animal. In unsupervised learning,
we don't provide the algorithm with labeled examples, so it has to figure out the
patterns on its own. This is like letting a child explore a zoo and try to figure out
which animals belong to which groups.

Machine learning is used in a wide variety of applications, from spam filtering to

image recognition to fraud detection. It's a powerful tool that can be used to solve a
wide range of problems, and it's only going to become more important in the years
to come.

3
What is Machine Learning?

Let us start with an example.

Let’s say you want to program a computer to recognize cats in pictures, you would
need to write code that explicitly defines all of the features of a cat. This would be
a very difficult task, as there are many different features that make up a cat, such as
its fur, eyes, ears, and nose.

Machine Learning (ML) algorithms can learn to recognize cats in pictures without
being explicitly programmed. They do this by training on a dataset of images that
are labeled as either containing a cat or not containing a cat. The algorithm then
looks for patterns in the images that are correlated with the labels. Once the
algorithm has learned these patterns, it can be used to predict whether or not a new
image contains a cat.

Definition of ML

Machine learning is a field of study (computer science) that enables computers to

learn from data without being explicitly programmed. Instead of explicitly coding

4
rules, ML algorithms can analyze data and identify patterns to make predictions or
decisions.

Types of Machine Learning

There are three/four main types of machine learning:

●​ Supervised learning:
In supervised learning, the computer is given a dataset of labeled
examples. The goal is to train the computer to learn the relationship between
the inputs (features) and the outputs (labels) so that it can make accurate
predictions for new, unseen data. For instance, a supervised learning
algorithm could be trained to classify images of cats and dogs based on a
dataset of labeled images.

Example 1: Predicting house prices

Data: Features of houses, such as size, location, and number of bedrooms
Labels: The selling price of each house

Features Labels

House ID Bedrooms Bathrooms Area (sq. ft.) Location Sale Price ($)

1 3 2 2000 Suburb A 300,000

2 4 3 2500 Urban Center 450,000

3 2 1 1500 Suburb B 240,000

... ... ... ... ... ...

Example 2: Diagnosing medical conditions

Data: Medical images, such as X-rays and MRIs

5
 Labels: Whether each image shows signs of a particular medical condition

●​ Unsupervised learning:
In unsupervised learning, the computer is given a dataset of unlabeled
data. The goal is to find patterns or structure in the data without being told
what to look for.

Example: Suppose you are given the following set of images. And you are
to cluster them into two groups. How would you group them?

Group by Gender:

Group 1

6
 Group 2

Well that is one possible feature to group them. How about this grouping?

Group 1​ ​ ​ ​ ​ ​ ​ Group 2

●​

7
●​ Self-supervised learning: Self-supervised learning bridges the gap between
supervised and unsupervised learning by generating labels or supervisory
signals from the data itself. This method has gained traction in recent years
due to its ability to handle large amounts of unlabeled data without relying
on expensive manual annotations.

Self-supervised learning obtains supervisory signals from the data itself,

often leve
raging the underlying structure in the data. The general technique of
self-supervised learning is to predict any unobserved or hidden part (or
property) of the input from any observed or unhidden part of the input. For
example, as is common in NLP, we can hide part of a sentence and predict
the hidden words from the remaining words. We can also predict past or
future frames in a video (hidden data) from current ones (observed data).
Since self-supervised learning uses the structure of the data itself, it can
make use of a variety of supervisory signals across co-occurring modalities
(e.g., video and audio) and across large data sets — all without relying on
labels.

8
 In self-supervised learning, the system is trained to predict hidden parts of
the input (in gray) from visible parts of the input (in green).
As a result of the supervisory signals that inform self-supervised learning,
the term “self-supervised learning” is more accepted than the previously
used term “unsupervised learning.” Unsupervised learning is an ill-defined
and misleading term that suggests that the learning uses no supervision at all.
In fact, self-supervised learning is not unsupervised, as it uses far more
feedback signals than standard supervised and reinforcement learning
methods do.

●​ Reinforcement learning:
In reinforcement learning, the computer learns through trial and error
by interacting with an environment. The goal is to maximize the reward
signal received by the computer. This is much like how you can train a dog
to do tricks. For example, a reinforcement learning algorithm (usually a deep
neural network) could be used to train a robot to play a game by rewarding it
for making successful moves.

Example 1: Training a self-driving car: Reinforcement learning algorithm

is used to train a self-driving car to navigate a road environment. The
algorithm learns to make decisions about acceleration, braking, and steering
based on the feedback it receives from the car's sensors.

Example 2: Playing complex games: Reinforcement learning algorithms

have been used to achieve superhuman performance in complex games such
as Go, Dota 2, and StarCraft-2. These algorithms are able to learn complex
strategies and adapt to their opponents in real-time.

Example 3: In chatbots: Reinforcement learning plays an important part in

training AI chatbots so that they do respond the way we want them to. This
is because we want to in some way tell the chatbots that what it is producing
is “good”. It has been observed that without human provided reinforcements
these chatbots behave quite erratically and can pick up pretty toxic
behavioral traits.

9
 In reinforcement learning the issue sometimes is that we set up the rewards
wrong, do not know how to set up the rewards, or the machine finds
unexpected ways to maximize its rewards and does something that was not
what it was designed to do. Many new algorithms are developed very
recently to tackle such problems. This is one of the most exciting areas in
machine learning.

Applications of Machine Learning

Machine learning is used in a wide range of applications, including:

●​ Image recognition: Identifying objects in images, such as faces, cars, and

animals
●​ Natural language processing (NLP): Understanding and generating human
language, such as translating languages and answering questions
●​ Recommendation systems: Recommending products, movies, or music to
users based on their preferences
●​ Fraud detection: Identifying fraudulent transactions in credit card or
insurance data
●​ Medical diagnosis: Assisting doctors in diagnosing diseases by analyzing
medical images and patient data
●​ Finance: AI in finance can help in five general areas: personalize services
and products, create opportunities, manage risk and fraud, enable
transparency and compliance, and automate operations and reduce costs.
●​ Drug Discovery: AI tools have the potential to transform drug discovery by
enabling researchers to rapidly analyze large-scale data sets, design new
molecules, and predict the efficacy of potential drug candidates. AI can help
in quick discovery of vaccines by using data and algorithms to simulate the
immune system, predict potential targets, design novel molecules, and
optimize clinical trials.

10
Discussion Points and Questions

Application Scenarios:

1.​ Crop Disease Identification:

Scenario: Developing a system to identify diseases in crops based on images
of leaves.
Question: Can this task of identifying diseases in crops be addressed using
supervised, unsupervised, or reinforcement learning?

2.​ Social Media Post Classification:

Scenario: Building a system to classify social media posts into relevant
categories (e.g., sports, news, entertainment).
Question: Which learning approach—supervised, unsupervised, or
reinforcement—might be suitable for categorizing these posts?
3.​ Customer Segmentation for Marketing:
Scenario: Dividing customers into different groups based on their purchasing
behavior and demographics for targeted marketing.
Question: How could supervised, unsupervised, or reinforcement learning
help in segmenting customers effectively?

4.​ Autonomous Driving:

Scenario: Creating a system for a self-driving car to navigate through city
streets and make driving decisions.
Question: Which learning method—supervised, unsupervised, or
reinforcement—could be utilized to train the car to drive safely?

5.​ Anomaly Detection in Network Traffic:

11
 Scenario: Detecting unusual patterns in network traffic to identify potential
cybersecurity threats.
Question: Is supervised, unsupervised, or reinforcement learning better
suited for detecting anomalies in network data?

More Discussion Points:

1.​ Crop Disease Identification:

Can we train a model using labeled images of diseased and healthy crops to
classify diseases (supervised)?
Is there an approach to identifying patterns in the images without explicit
labels (unsupervised)?

2.​ Social Media Post Classification:

Can we use historical data with labeled categories to train a classifier
(supervised)?
Is there a way to categorize posts without predefined labels or categories
(unsupervised)?

3.​ Customer Segmentation for Marketing:

Can we use labeled data (customer behavior with predefined segments) to
create a segmentation model (supervised)?
Can the model identify natural groupings or patterns within customer data
without predefined segments (unsupervised)?

4.​ Autonomous Driving:

Can the car learn safe driving behaviors through trial and error
(reinforcement)?
Can it learn from examples of safe driving behaviors (supervised)?

5.​ Anomaly Detection in Network Traffic:

Can we use historical data labeled as normal/abnormal to train a detection
model (supervised)?

12
Can the model identify abnormal patterns in the absence of labeled
anomalies (unsupervised)?

13
Machine Learning Models [Optional]

K-nearest neighbors (KNN)

KNN is a simple, yet powerful, machine learning algorithm that can be used for
both classification and regression tasks. It is based on the idea that similar data
points are likely to have similar outcomes.

- Class A (Dog)​
- Class B (Cat)

Look at this image, there are two different data points, let’s say they represent two
different classes, class A or red triangles represent images of dogs and class B, blue
squares represent images of cats. But don’t know what the green circle represents,
we have to classify this green circle, is it a cat or a dog?

14
Decision Trees

Below is an example of a machine learning model that is called a decision tree. The
box in the top of the diagram labeled “Age” is called the root of the tree and the
boxes labeled “Granted” or “not Granted” are the leaves. A decision tree is more
like a tree drawn upside down.

This particular decision tree is a simplified model that illustrates how a bank might
use a person's age, income, average monthly savings amount, and credit duration to
decide whether or not to grant them a loan. A set of data like the one below with
the actual loan granting decision made is used in the decision tree.
.
Age Income Average Monthly Credit Decision
Savings Amount Duration

38 170000 675 35 Granted

43 165000 430 40 Not Granted
40 180000 700 30 Granted
42 162000 420 40 Not Granted

15
35 160000 630 40 Not Granted
… … … … …

See if you can figure out what will happen if the decision tree is asked to make a
decision on the following data:

Age Income Average Monthly Credit Decision

Savings Amount Duration

38 162000 650 37 ?

Artificial Neural Networks

Artificial Neural Networks are networks of Artificial Neurons which are basically
mathematical models of neurons in the brain. These models have become one of
the most popular and useful models in recent years.

Biological Neurons:
In our brains, we have billions of tiny cells called neurons. These neurons receive
signals from other neurons through connections called synapses. When a neuron
gets enough signals, it "fires" an electrical signal down its own length, passing the
message to the next neuron in the chain.

Artificial Neurons
Artificial neurons are like simplified versions of these biological neurons. They
receive inputs (like the signals in our brains), each with a different importance (like
the strength of the synapses). The artificial neuron then adds up these weighted
inputs and decides whether to "fire" or not, just like a biological neuron.

16
 Figure: Biological neuron and artificial neuron

Artificial neurons have three main components:

●​ Inputs: The inputs to an artificial neuron are the values from other artificial
neurons or from external data sources. These are x1, …, xn in the diagram
above.
●​ Weights: The weights of an artificial neuron determine how much influence
each input has on the neuron's output. These weights are w1, w2, …, wn.
You can also see x0=1 and a bias b in the diagram above. These are used to
compute the weighted inputs (Z) in the following way:
●​ Z = x1*w1 + x2*w2 + … + xn*wn + b
●​ Activation function: The activation function of an artificial neuron
determines the neuron's output (y) based on the weighted inputs (Z). An
example of activation function is the step function
○​ y = step(z): if z > 0 then y = 1, otherwise y = 0.

17
Check out this artificial neuron. This is supposed to compute the logical AND of
the two binary inputs x1 and x2. Examine for each input pair, how it computes the
desired output.

Can you create an artificial neuron that can compute OR and the NOT operations?
Can you combine them in some way to construct networks of artificial neurons that
can compute NAND and XOR operations of the inputs.

Biological Neural Networks

Our brains have complex networks of these neurons working together. Each
neuron's firing influences others, and this interconnected web of neurons allows
our brains to process information, learn, and make decisions.

Artificial Neural Networks

Similarly, artificial neural networks consist of layers of artificial neurons connected
to each other. The connections have weights (like the strength of synapses), and the
network learns by adjusting these weights based on its successes and mistakes. The
network's structure and the way neurons work together help it solve tasks, like
recognizing images or understanding language.

18
Neural networks are typically organized into layers. The first layer of a neural
network is called the input layer. Each neuron in the input layer receives a single
input value. The last layer of a neural network is called the output layer. Each
neuron in the output layer produces a single output value or multiple outputs.

The layers between the input layer and the output layer are called hidden layers.
The hidden layers are responsible for learning the complex relationships between
the data and the output. Learning algorithms set the weights of each neuron in a
way so that the output layer produces the desired output (or something close to it)
for a particular input.

So, while artificial neurons and neural networks are simplified versions of their
biological counterparts, they borrow the concept of interconnected decision-makers
from our brains to perform tasks and learn from experience.

19
Deep Learning

Deep learning is a type of machine learning that involves using artificial neural
networks with many layers, often referred to as deep neural networks. The term
"deep" comes from the idea of having multiple layers (deep architectures) in the
network.

Deep learning models are typically very large, with millions or even billions of
parameters. This allows them to learn very complex patterns from data, and to
achieve state-of-the-art performance on a wide range of tasks. For example, deep
learning models have been used to achieve superhuman performance on tasks such
as image recognition, natural language processing, and speech recognition.

Here are some of the most popular types of deep learning models:

●​ Convolutional neural networks (CNNs) are a type of ANN that is

well-suited for image recognition tasks. CNNs are able to learn complex
features from images, such as edges, shapes, and textures.

Typical CNN model diagram

●​ Recurrent neural networks (RNNs) are a type of ANN that is well-suited

for natural language processing tasks. RNNs are able to learn the sequential
nature of language, which is important for tasks such as machine translation
and text summarization.

20
 ●​ Transformer architecture is another significant development in the realm of
deep learning, especially in natural language processing (NLP). Transformer
models are based on the principle of self-attention, which allows them to
learn long-range dependencies between words in a sentence.

These models typically try to represent each input image, text data (words,
sentences, documents) or audio as a list of numbers (vectors) so that similar
images, text data or audio have similar representations. These representations are
called embeddings. This becomes useful when we try to then use neural networks
to classify inputs or reproduce similar inputs. Below is an example visualization of
book embeddings. Embeddings themselves are not two dimensional but are
projected into two dimensions (tSNE is a type of visualization technique) for
visualization. Here we can see that embeddings of books of the same genre are
clustered together quite nicely.

21
Deep learning models are still a relatively new area of research, but they have
already had a significant impact on the field of artificial intelligence. Deep learning
models are now being used in a wide range of applications, including:

Generative Models​

A generative model is a type of machine learning model that aims to learn the
underlying patterns or distributions of data in order to generate new, similar data.
In essence, it's like teaching a computer to dream up its own data based on what it
has seen before. The significance of this model lies in its ability to create, which
has vast implications in various fields, from art to science. Examples include
LLMs, Diffusion based image generators.

LLM based Model: ChatGPT

Prompt: Write a poem about autumn evening.

ChatGPT:
As daylight wanes, a crimson sky unfurls,
Autumn whispers secrets to the world,
Leaves rustle, painting paths in gold,
A crisp embrace, the air's story told,
Twilight's hush in the evening's reign,
A fleeting dance before night's domain.

Asked to translate the first paragraph of this section this is what it produced

22
Prompt: You are an expert translator. Your task is to translate the following text to
Bangla. Please make sure the translated text is natural, easy to follow, and has
literary quality to it.
[...text…]

ChatGPT:

একটি জেনারেটিভ মডেল হল মেশিন লার্নিং মডেলের এক ধরন, যা তথ্যের নিচে লুকিয়ে থাকা প্যাটার্ন বা তথ্যের
বিতর্কি ত বিতর্কি ত ভাগবিভাজন শেখার লক্ষ্যে তার সামনে নতু ন, অনুরূপ তথ্য তৈরি করা। মূলত, এটি হলো যে, যেটি
কোম্পিউটারকে শিখানোর মতো, যাতে তা আগে দেখা তথ্য ভিত্তিক নিজের তথ্য সৃষ্টি করতে পারে। এই মডেলের গুরুত্ব
সেই সৃষ্টিশীলতায় অবদান, যা বিভিন্ন ক্ষেত্রে, শিল্প থেকে বিজ্ঞানে বিশাল প্রভাব ফেলে।

Not quite up to the mark is it? what do you think?

Diffusion Based Model: DALL-E​

​
Prompt: A boy flying a kite in the paddy field, in the summer afternoon, create the
art in the style of Van Gaugh.
​
DALL-E:

However, these models can produce incoherent images too. See 5 things AI image
generators still struggle with | Digital Trends

Benefits of Machine Learning

Machine learning offers several benefits, including:

23
●​ Improved decision-making: ML algorithms can analyze large amounts of
data and identify patterns that humans might miss, leading to better
decisions.
●​ Automation: ML can automate tasks that are repetitive or time-consuming,
freeing up human workers to focus on more creative and strategic work.
●​ Personalization: ML can personalize products, services, and
recommendations to individual users, providing a more tailored experience.
●​ Innovation: ML is driving innovation in a wide range of industries, leading
to new products, services, and processes.

24
Distinguishing the terms ML, DL & AI

Artificial Intelligence (AI) is a broad term that refers to the ability of machines to
perform tasks that are typically associated with human intelligence, such as
learning, reasoning, and problem-solving.

Machine Learning (ML) is a subset of AI that deals with the development of

algorithms that can learn from data without being explicitly programmed. ML
algorithms are trained on a set of data, and they can then use that data to make
predictions or decisions.

Deep Learning (DL) is a subset of ML that uses artificial neural networks to learn
from data. Neural networks are inspired by the human brain, and they are able to
learn complex patterns from data.

25
Why is AI a big deal?

AI is a big deal because it is having a profound impact on our world in many

different ways. Here are some of the reasons why AI is so important:

●​ AI is driving innovation and economic growth. AI is being used to develop

new products and services, automate tasks, and improve efficiency. This is
leading to increased productivity, economic growth, and new job
opportunities.
●​ AI is improving decision-making. AI can analyze large amounts of data and
identify patterns that humans would miss. This can help businesses make
better decisions about everything from product development to marketing.
●​ AI is making our lives easier. AI is being used to automate tasks such as
driving cars, translating languages, and answering our questions. This is
freeing up our time to focus on more important things.
●​ AI is changing the way we interact with the world around us. AI is being
used to develop new forms of entertainment, education, and communication.
This is changing the way we learn, work, and play.

Why now?

Artificial intelligence is a big deal now because a couple of technologies:

●​ The increase in computing processing power, which allows AI systems to

handle large amounts of data and complex calculations faster and more
efficiently.
○​ NVIDIA A100, Google TPU v3, 4 are some of the popular GPUs in
the field of AI. And because of this exceptional performance AI can
produce remarkable results, more on this later (why AI is a big deal
now).

26
 ○​ NVIDIA A100 delivers 312 teraFLOPS (trillion floating point
operation per seconds) of deep learning performance.
●​ The rise of big data, which provides AI systems with more information and
examples to learn from and discover patterns and insights.
○​ Big data is a large and complex collection of data that is constantly
growing. It can be structured, unstructured, or semi-structured. Big
data is generated and transmitted from many sources.
○​ Sources of big data, including, social media, IoT systems, Banking,
Audio Video, Government and Public Data, Research and Scientific
data, and so on.
●​ The adoption of cloud-based computing, which enables AI systems to access
and share data and resources across different platforms and devices.
○​ Cloud computing is the on-demand availability of computing
resources (such as storage and infrastructure), as services over the
internet. It eliminates the need for individuals and businesses to
self-manage physical resources themselves, and only pay for what
they use.
○​ Some cloud computing platforms – Amazon web services (AWS),
Microsoft Azure, Google Cloud Platform (GCP), etc.

27
Ethical issues and other Problems with AI

On the other hand, AI poses more real dangers. Here we discuss a couple of them:

1.​ AI might create loss of jobs and more inequality

AI, led by large companies, holds immense potential but also raises concerns about
intensifying inequalities. It threatens widespread job displacement, with estimates
suggesting up to 800 million jobs could be automated by 2030. Moreover, the
advantage lies with large companies equipped to develop and deploy superior AI
systems, further widening the wealth gap. To counter these effects, retraining and
reskilling programs are crucial to help displaced workers adapt to the AI-driven job
market. Governments, businesses, and educational institutions must collaborate:
governments can fund training programs, businesses can provide apprenticeships,
and educational institutions can adapt curricula for the AI era.

2.​ AI systems can be biased

Bias and discrimination can be introduced at various stages, from data collection
and preparation to modeling and analysis, and even interpretation and
communication. Biased datasets in the context of AI refer to datasets that do not
adequately represent the diversity of the real-world population, leading to skewed
or unfair outcomes when machine learning models are trained on them. For
example, consider facial recognition technology trained predominantly on images
of lighter-skinned individuals. When faced with faces of darker skin tones, this AI
system might struggle.

3.​ Introduces complicated copyright issues

AI data ownership is a complex matter influenced by data type, AI system,

jurisdiction, and contracts. Generally, data lacks explicit intellectual property rights
but may have legal protections (privacy, trade secrets). Ownership often ties to data
origin and permissions granted to the AI system. The generation of new data by AI
raises questions about ownership. Lawsuits between artists and AI companies

28
underscore uncertainties—highlighting the evolving nature of data ownership in
AI. Clear data governance, contracts, and legal compliance offer potential solutions
in navigating this intricate issue.

4.​ Current AI models are black boxes

A black box model is a type of model that is difficult or impossible to understand

how it works. This is because the model is trained on a large amount of data, and
the relationships between the inputs and outputs of the model are complex and
nonlinear. As a result, it is difficult for humans to interpret the model's
decision-making process. Explainable AI (XAI) is a field of research that aims to
make black box models more transparent.

5.​ Who is responsible for mistakes made by AI systems?

AI accountability is a complex issue with multiple stakeholders involved. When AI

fails, it can be difficult to determine who is responsible. Potential responsible
parties include the AI developer, the data supplier, the organization deploying the
AI, regulatory bodies, and even the end user. To ensure AI accountability, we need
robust regulations, transparent practices, and a commitment to continuous learning
and adaptation.

6.​ Harmful content creation using AI

AI-generated harmful content can take various forms, spanning from

misinformation and hate speech to deep fakes. These forms of content can have
far-reaching consequences, both for individuals and for society as a whole. To
combat AI-generated harmful content, we need to focus on responsible
development, regulation, and ethical use of AI technologies. We also need to
educate the public about the risks of harmful content and provide them with tools
to identify and avoid it.

7.​ The value alignment problem

29
AI alignment is the process of ensuring that AI systems are aligned with human
values and goals. It is important because AI systems are becoming increasingly
autonomous and are making decisions that have a significant impact on our lives.
Without AI alignment, there is a risk that AI systems could act in ways that are
harmful to humans.

8.​ AI singularity?

The AI singularity is a hypothetical future event in which AI surpasses human

intelligence, leading to rapid technological advancements. This could bring
significant societal and technological changes. While some argue it would lead to
rapid progress and a deeper understanding of the universe, others fear AI becoming
uncontrollable or replacing humans. The singularity is uncertain, and we should
consider all perspectives and be prepared for the future.

Building safe AI systems is crucial as AI permeates various aspects of our lives.

For example, The DeepMind Safety Research team has outlined a framework to
build safe AI systems: emphasizing clearly defined objectives, addressing biases,
designing for robustness, ensuring transparency, establishing oversight, continuous
monitoring, public engagement, and international cooperation. Additionally,
technical challenges such as verifying algorithm correctness, preventing
unintended consequences, and ensuring long-term safety need to be addressed.
Building safe AI is an ongoing process that requires collaboration, innovation, and
adaptation.

AI is not a static or fixed technology, but a dynamic and evolving one. AI is

constantly changing and improving, as researchers and developers discover new
methods, algorithms, and applications. AI is also influenced by the feedback and
interaction of its users and stakeholders, who shape its goals, values, and outcomes.
Therefore, AI is not only a product of human ingenuity, but also a reflection of
human society.

30
Discussion Points

●​ Which jobs are more likely to be replaced by AI?

●​ Which demands for which jobs are likely to grow?

●​ What new jobs can AI create?

●​ Who should be responsible for making AI systems safe, unbiased?

31
How to build AI systems?

Building an AI system usually involves several key steps:

●​ defining the problem
●​ gathering relevant data
●​ selecting the right AI technique (like machine learning or deep learning)
●​ preparing and training the model
●​ evaluating its performance, and
●​ deploying it while continuously monitoring for improvements.

AI APIs provide ready-made AI functionalities, simplifying development. To use

them effectively: understand your needs, choose the appropriate API (e.g., Google
Cloud Vision for image recognition), integrate it following guidelines, test
thoroughly, and maintain it post-deployment.

Skills required for AI development: skills span a number of different fields

●​ programming (some popular languages for AI/ML are Python, R)
●​ mathematics (linear algebra, calculus, statistics)
●​ machine learning concepts
●​ Data Engineering skills: data handling (cleaning, transforming), data
pipeline development
●​ DevOPs: these are the skill required to deploy large AI systems
●​ Understanding of the ethical considerations of the project, and
●​ ongoing learning through courses, projects, and staying updated with the
latest advancements.

32