Introduction to AI

Dirk deRoos
WW Technology Sales Enablement
Data & AI Manager

Artificial intelligence (AI) leverages computers and machines attempt to mimic (at scale) the problem-solving and
decision-making capabilities of the human mind. AI systems are “trained and learned” from experience through an
iterative process. AI technology can remember behavior patterns and adapt their responses to conform to those
behaviors or encourage changes.

Currently, AI is being used in everyday life across many domains, from drastically improving travel with
applications like Google and Apple Maps, to using face ID to unlock a phone or application, to even predicting the
kinds of food people will order on a hot summer day to power a dynamically-priced drive-thru fast food menu.
Smart recommendation systems have learned consumers behaviors and interests over time by following online
activity and use AI to align suggestions with an individual’s interests and buying behaviors (referred to as hyper-
personalization). Social media applications are using AI to monitor content for hateful speech or inappropriate
images, suggest new connections within a social network, and serve advertisements to targeted users for revenue.
AI is playing a significant role in other industries such as banking (risk, upsell) and healthcare (reduced
readmissions) as well and AI technology can help organizations find answers to questions, they may not even
know they should be asking.

The truth is… there isn’t an industry where AI isn’t being put to work, or (more importantly) couldn’t be put to work
in, because when done right, AI can help companies optimize, predict, or automate their business.

This presentation is an introduction to AI concepts including machine learning methods, foundation models, and
traditional AI applications and use cases. This presentation helps learners prepare for conversations about AI
concepts and use cases.

Notes:

1
• Machine learning (ML) refers to a broad set of techniques to train a computer to learn from
its inputs, using existing data, and one or more “training” methods, instead of being
explicitly programmed. ML helps a computer to achieve AI.

• Foundation models are typically built using a specific kind of neural network architecture,
called a transformer, which is designed to generate sequences of related data elements (for
example, like a sentence).

• Neural networks are a set of algorithms, modeled loosely after the neural networks found
in the human brain, that are designed to recognize hidden patterns in data.

• A transformer model is a neural network architecture useful for understanding language,

which does not have to understand words one at a time but can look at an entire sentence
at once for context and disambiguation. A technique called self-attention allows these
models to focus on the particular words that are important in understanding the meaning of
the sentence. Transformers single-handedly spawned the world of foundation models and
its birthplace research paper (Attention is all You Need) spawned the AI inflection point
that is happening today. AI transformer models can be Encoder models (like BERT),
Decoder models (like GPT), or Encoder and Decoder models (like T5). Learn more about
Transformers at: https://www.youtube.com/watch?v=ZXiruGOCn9s.

• An algorithm is a procedure used for solving a problem or performing a computation.

Algorithms act as an exact list of instructions that conduct a sequence of specified actions
in either hardware- or software-based routines.

1
 Legal disclaimer © IBM Corporation 2023.
All Rights Reserved.
References in this presentation to IBM
products, programs, or services do not
imply that they will be available in all
The information contained in this countries in which IBM operates. Product
publication is provided for informational release dates and/or capabilities referenced
purposes only. While efforts were made in this presentation may change at any time
to verify the completeness and accuracy at IBM’s sole discretion based on market
of the information contained in this opportunities or other factors and are not
publication, it is provided AS IS without intended to be a commitment to future
warranty of any kind, express or implied. product or feature availability in any way.
IBM shall not be responsible for
any damages arising out of the use of, or All client examples described are presented
otherwise related to, this publication or as illustrations of how those clients have
any other materials. Nothing contained in used IBM products and the results, they
this publication is intended to, nor shall may have achieved. Actual environmental
have the effect of, creating any warranties costs and performance characteristics
or representations from IBM or its suppliers may vary by client.
or licensors, or altering the terms and
conditions of the applicable license
agreement governing the use of
IBM software.

© Copyright IBM Corp. 2023

This self-explanatory slide shares legal disclaimers for this presentation.

2
 What is AI?
AI refers to the ability
of computer systems
to attempt to mimic
the problem-solving
and decision-making Computer vision Data science
capabilities of the
human mind.

Natural language Robotics

processing (NLP)

Artificial intelligence (AI) is a general-purpose term which does not have a precise technical meaning. More
broadly, AI refers to the ability of computer systems to attempt to mimic the problem-solving and decision-making
capabilities of the human mind. AI is often used to refer to many disciplines, like natural language processing
(NLP), computer vision, robotics, data science, and more.

At a high level, AI is also considered to be a domain that combines computer science and robust datasets, that
enables problem-solving. This usage of the AI term encompasses the subfields of machine learning (ML), deep
learning (DL), and foundation models. These disciplines are comprised of algorithms that make predictions,
classifications, or generate content based on input data.

Notes:

• Natural language processing (NLP) is the technology that gives computers the ability to understand text and
spoken words in much the same way human beings can. NLP combines computational linguistics — rule-based
modeling of human language — with statistical, machine learning (ML), and deep learning (DL) models. These
technologies enable computers to process human language in the form of text or voice data and to ‘understand’
its full meaning, complete with the speaker or writer’s intent and sentiment.

• Neural networks are a set of algorithms, modeled loosely after the neural networks found in the human brain,
that are designed to recognize hidden patterns in data.

• An algorithm is a procedure used for solving a problem or performing a computation. Algorithms act as an
exact list of instructions that conduct a sequence of specified actions in either hardware- or software-based
routines.

3
• Deep learning (DL) is a technique for implementing machine learning (ML) that relies on
deep artificial neural networks to perform complex tasks such as image recognition, object
detection, and natural language processing (NLP). Neural networks are a set of algorithms,
modeled loosely after the neural networks found in the human brain, which are designed to
recognize hidden patterns in data. The “deep” in DL refers to a neural network comprised of
more than three layers (which include the input and the output layers).

• Machine learning (ML) refers to a broad set of techniques to train a computer to learn from
its inputs, using existing data, and one or more “training” methods, instead of being
explicitly programmed. ML helps a computer to achieve AI.

• Foundation models are typically built using a specific kind of neural network architecture,
called a transformer, which is designed to generate sequences of related data elements (for
example, like a sentence).

• A transformer model is a neural network architecture useful for understanding language,

which does not have to understand words one at a time but can look at an entire sentence
at once for context and disambiguation. A technique called self-attention allows these
models to focus on the particular words that are important in understanding the meaning of
the sentence. Transformers single-handedly spawned the world of foundation models and
its birthplace research paper (Attention is all You Need) spawned the AI inflection point
that is happening today. AI transformer models can be Encoder models (like BERT),
Decoder models (like GPT), or Encoder and Decoder models (like T5). Learn more about
Transformers at: https://www.youtube.com/watch?v=ZXiruGOCn9s.

3
 AI milestones

1956 1980s 2011 2022

John McCarthy coins Neural networks IBM Watson beats OpenAI releases
the term 'artificial become widely used champions Ken ChatGPT, a generative
intelligence' at the first- in AI applications. Jennings and Brad AI chatbot based on a
ever AI conference at Rutter at Jeopardy! large language model.
Dartmouth College.

1950 1967 1997 2016

Alan Turing published the Frank Rosenblatt builds IBM's Deep Blue DeepMind's AlphaGo
Computing Machinery the Mark 1 Perceptron, supercomputer beats, program, powered
and Intelligence paper the first computer based then world chess by a deep neural
introducing the question: on a neural network champion, Garry network, beats Lee
“Can machines think?” that ‘learned’ through Kasparov in a chess
Sedol, the world
trial and error. match (and rematch).
champion Go player,
in a five-game match.

Artificial intelligence (AI) is by no means a new concept. The birth of the AI conversation was denoted by Alan
Turing, often referred to as the “father of computer science,” in the paper "Computing Machinery and Intelligence"
published in 1950. In that paper, Turing asks the following question, "Can machines think?" He offers a test, now
famously known as the Turing test. The Turing test challenges a human interrogator to try and distinguish an
interaction between a computer and human in a text response. While this test has undergone much scrutiny since
published, it remains an important part of the history of AI.

In 1956, John McCarthy coins the term ‘artificial intelligence’ at the first-ever AI conference at Dartmouth College.
Much later, in his 2004 “What is Artificial Intelligence?” paper, McCarthy offers this definition of AI: “It is the
science and engineering of making intelligent machines, especially intelligent computer programs. It is related to
the similar task of using computers to understand human intelligence, but AI does not have to confine itself to
methods that are biologically observable.“

In 1967, Frank Rosenblatt builds the Mark 1 Perceptron, the first AI based on a neural network that ‘learned’
through trial and error. A year later, Marvin Minsky and Seymour Papert publish a book titled “Perceptrons.” This
book becomes the landmark work of neural networks… by the 1980s, neural networks become widely used in AI
applications.

In 1997, IBM's Deep Blue supercomputer beats then world chess champion, Garry Kasparov, in a chess match
(and rematch). It would be just over a decade later when IBM Watson competes and beats champions Ken
Jennings and Brad Rutter at Jeopardy! in 2011.

Supercomputers continue to train models that advance AI. In 2015, the Baidu Minwa supercomputer uses a
special kind of deep neural network, called a convolutional neural network (CNN), to identify and categorize
images with a higher rate of accuracy than the average human. And in 2016, the DeepMind AlphaGo program,

4
powered by a deep neural network, beats Lee Sedol, the world champion Go player, in a five-
game match. (Go is an adversarial game with the objective of surrounding a larger total area of
the board with one's stones than the opponent. As the game progresses, players position
stones on the board to map out formations and potential territories. Go is considered the most
difficult game ever invented by man). This victory is significant given the huge number of
possible moves as the game progresses (over 14.5 trillion after just four moves!).

In 2022, the OpenAI research lab released ChatGPT, a public prototype of a generative AI
system family (Generative Pre-trained Transformers… GPT for short) that automatically
creates text based on written prompts in a manner that is highly advanced, extremely flexible,
but also conversational in nature. ChatGPT is based on a large language model (LLM) (at least
initially) called GPT-3.5 (and now GPT-4 for subscribers to ChatGPT+ (over time these
versions will change)). ChatGPT’s LLM that was trained on a large body of text and code from a
variety of sources (like Wikipedia, 3-star or above ratings on Reddit reviews, and more). Once it
answers a question, ChatGPT can amend, reword, reorganize, or even rewrite it in a different
language or style through subsequent interactions. The release of ChatGPT (and other AIs —
like Stable Diffusion with its ability to make images) has led to unprecedented widespread
interest in AI.

AI is an ever-evolving sector of computer science that focuses on smart technology —

machines that can perform tasks that would otherwise require human intelligence or intuition.
It has garnered both criticism and praise, with some warning that AI may change the
workforce in a negative way forever, while others suggest that it can revolutionize healthcare
and other industries. Regardless, AI is on a trajectory to becoming more integrated into society
than ever imagined.

Notes:

• Neural networks are a set of algorithms, modeled loosely after the neural networks found
in the human brain, that are designed to recognize hidden patterns in data.

• An algorithm is a procedure used for solving a problem or performing a computation.

Algorithms act as an exact list of instructions that conduct a sequence of specified actions
in either hardware- or software-based routines.

• Convolutional Neural Network (CNN) is a network architecture for deep learning

specifically used for image recognition and tasks that involve the processing of pixel data.
CNNs are particularly useful for finding patterns in images to recognize objects, classes,
segments (different ‘things’ in a picture), and categories.

• Deep learning (DL) is a technique for implementing machine learning (ML) that relies on
deep artificial neural networks to perform complex tasks such as image recognition, object
detection, and natural language processing (NLP). Neural networks are a set of algorithms,
modeled loosely after the neural networks found in the human brain, which are designed to
recognize hidden patterns in data. The “deep” in DL refers to a neural network comprised of
more than three layers (which include the input and the output layers).

4
• Go is a board game that originated in China over 3,000 years ago. Winning this board game
requires multiple layers of strategic thinking.

• GPT stands for Generative Pre-trained Transformer and refers to a family of neural
network-based large language models (LLMs) developed by OpenAI; for example, GPT-3
and GPT-4.

• Generative AI refers to a set of AI algorithms that can generate new outputs — such as
text, images, code, or audio — based on the training data, unlike traditional AI systems that
are designed to recognize patterns and make predictions. Sometimes the AI that powers
these solutions are referred to as decoders.

• A large language model (LLM) is a type of machine learning model that has been trained
on large quantities of unlabeled text using self-supervised learning and can perform a
variety of natural language processing (NLP) tasks (even when that language is a
programming language). Output may range from books, articles, social media posts, online
conversations, and even code. The architecture of an LLM consists of layers of neural
networks that learn to generate language in a way that is similar to how humans use
language.

• Natural language processing (NLP) is the technology that gives computers the ability to
understand text and spoken words in much the same way human beings can. NLP
combines computational linguistics — rule-based modeling of human language — with
statistical, machine learning (ML), and deep learning (DL) models. These technologies
enable computers to process human language in the form of text or voice data and to
‘understand’ its full meaning, complete with the speaker or writer’s intent and sentiment.

• A transformer model is a neural network architecture useful for understanding language,

which does not have to understand words one at a time but can look at an entire sentence
at once for context and disambiguation. A technique called self-attention allows these
models to focus on the particular words that are important in understanding the meaning of
the sentence. Transformers single-handedly spawned the world of foundation models and
its birthplace research paper (Attention is all You Need) spawned the AI inflection point
that is happening today. AI transformer models can be Encoder models (like BERT),
Decoder models (like GPT), or Encoder and Decoder models (like T5). Learn more about
Transformers at: https://www.youtube.com/watch?v=ZXiruGOCn9s.

• Foundation models are typically built using a specific kind of neural network architecture,
called a transformer, which is designed to generate sequences of related data elements (for
example, like a sentence).

Sources:

• What is artificial intelligence? Webpage: https://www.ibm.com/topics/artificial-intelligence

4
• Computing Machinery and Intelligence paper:
https://redirect.cs.umbc.edu/courses/471/papers/turing.pdf

• What is Artificial Intelligence? paper: https://www-formal.stanford.edu/jmc/whatisai.pdf

4
 Artificial Intelligence (AI) AI can be defined as a technique that enables machines to mimic
cognitive functions associated with human minds – cognitive functions
Human intelligence exhibited by machines
include all aspects of learning, reasoning, perceiving, and problem solving.

Machine Learning (ML) ML-based systems are trained on historical data to uncover patterns.
Users provide inputs to the ML system, which then applies these inputs
Systems that learn from historical data
to the discovered patterns and generates corresponding outputs.

Deep Learning (DL) DL is a subset of ML, using multiple layers of neural networks, which are
interconnected nodes, which work together to process information. DL is
ML technique that mimics
well suited to complex applications, like image and speech recognition.
human brain function

Foundation Model
Generative AI systems

AI model built using a specific kind of neural network architecture,

called a transformer, which is designed to generate sequences
of related data elements (for example, like a sentence).

1950’s 1980’s 2010’s 2020’s

Artificial intelligence (AI), machine learning (ML), and deep learning (DL) are three popular terms that are used,
often interchangeably, to describe how computers turn data into valuable insights. But, as this slide shows, these
terms are not interchangeable. Instead, you can think of AI, ML, and DL as a set of Russian dolls nested within
each other, with the largest, outermost doll being AI and the smallest, innermost doll being DL (leaving ML in the
middle).

AI is the all-encompassing concept that’s concerned with building smart machines that are capable of performing
tasks that normally require human intelligence. Machine learning refers to a broad set of techniques that give
computers the ability to “learn” by themselves, using existing data and one or more “training” algorithms, instead
of being explicitly programmed. And deep learning is a technique for implementing ML that relies on deep artificial
neural networks to perform complex tasks such as image recognition, object detection, and natural language
processing. (Neural networks are a set of algorithms, modeled loosely after the neural networks found in the
human brain, that are designed to recognize hidden patterns in data.)

Machine learning is a subset of AI. What separates it from rules-based engines is its ability to improve its
performance when exposed to more data. The “learning” part of ML means that algorithms (or models) attempt to
optimize along a certain dimension; typically, to minimize error or to maximize the likelihood their predictions will
be correct.

The concept of deep learning has been around for quite some time as well. Deep learning is a subset of ML and
normally, when people use the term deep learning, they are referring to deep artificial neural networks. In this
case, “deep” is a technical term that refers to the number of layers that are used in the neural network itself – a
shallow network will have one hidden layer while a deep network can contain hundreds. Multiple hidden layers
allow deep neural networks to learn features of data in a so-called feature hierarchy, where simple features (for
example, pixels) recombine from one layer to the next, to form more complex features (for instance, lines). Neural

5
networks with many layers pass data (features) through more mathematical operations than
those with fewer layers, and therefore are more computationally intensive to train.

Prior to 2010, very little work was being done in DL because even shallow neural networks are
extremely compute-intensive. It was only when three forces converged – the availability of
immense stores of labeled data; the invention of new deep learning algorithms; and cheaper,
more powerful CPUs (central processing units) coupled with the intense performance of GPUs
(graphics processing units designed to render images, animations, and video for a computer's
screen, which are exceptional at the matrix-matrix multiplication that is often utilized by deep
learning) – that deep learning moved from research papers and laboratories to real world
applications. Now, the way in which deep learning can be adopted and used has become a
critical strategic consideration for every cognitive enterprise.

As pointed out earlier, the terms AI, ML, and DL are frequently misused, but all three
technologies encompass the process of analyzing immense volumes of data to find common
patterns and then turning those patterns into actionable predictions and insights.

For more information, refer to this article which was the inspiration/source for much of this
narrative and the visual: https://blogs.nvidia.com/blog/2016/07/29/whats-difference-
artificial-intelligence-machine-learning-deep-learning-ai/.

Notes:

• Foundation models are typically built using a specific kind of neural network architecture,
called a transformer, which is designed to generate sequences of related data elements (for
example, like a sentence).

• Neural networks are a set of algorithms, modeled loosely after the neural networks found
in the human brain, that are designed to recognize hidden patterns in data.

• A transformer model is a neural network architecture useful for understanding language,

which does not have to understand words one at a time but can look at an entire sentence
at once for context and disambiguation. A technique called self-attention allows these
models to focus on the particular words that are important in understanding the meaning of
the sentence. Transformers single-handedly spawned the world of foundation models and
its birthplace research paper (Attention is all You Need) spawned the AI inflection point
that is happening today. AI transformer models can be Encoder models (like BERT),
Decoder models (like GPT), or Encoder and Decoder models (like T5). Learn more about
Transformers at: https://www.youtube.com/watch?v=ZXiruGOCn9s.

• An algorithm is a procedure used for solving a problem or performing a computation.

Algorithms act as an exact list of instructions that conduct a sequence of specified actions
in either hardware- or software-based routines.

• Natural language processing (NLP) is the technology that gives computers the ability to

5
understand text and spoken words in much the same way human beings can. NLP
combines computational linguistics — rule-based modeling of human language — with
statistical, machine learning (ML), and deep learning (DL) models. These technologies
enable computers to process human language in the form of text or voice data and to
‘understand’ its full meaning, complete with the speaker or writer’s intent and sentiment.

5
 Rules-based systems Machine learning systems

programmed with a series trained with a

of instructions & logical rules series of examples
for example: “a triangle has three sides” for example: “here are
pictures of many triangles”

Traditional applications are programmed using series of logical rules. For example, if someone was to write an
application to categorize different shapes, a programmer would define a set of rules capturing the characteristics
of those shapes; like, every triangle has three connected sides. For simple situations, it’s generally easy to program
applications where situations call for simple rules. However, for more complex situations, like programming a
chatbot identifying a type of car and so on, the application logic would be incredibly complex and difficult to
program and maintain.

In contrast to rules-based systems, intelligent systems (using artificial intelligence (AI) technology) are trained,
not programmed. Instead of spelling out specific rules to solve a problem, practitioners feed the AI system
examples of what it will encounter in the real world, and let the AI detect the patterns and produce its own ‘rules’.
Quite simply, the AI builds its own logic, and practitioners explain the domain of interest to the AI with labeled
examples instead of instructions.

The AI systems approach allowing machines to find patterns is beneficial over spelling out the instructions (the
rules-based systems approach)… especially when the instructions are hard or unknown or when the data has
many different variables; for example, developing medications, or predicting the stock market. What’s more, to
change the focus of the AI, feed the AI different data versus reprogram the rules… this allows for efficiency like
transfer lessons.

This leads to such benefit: portability. While AI systems need to be fine-tuned during the building phase, getting an
AI to recognize a cat instead of a dog only requires pictures of cats to be ‘fed’ to AI systems… not a redefinition of
pre-established rules. (In fact, using a process called transfer learning, an AI can quickly learn the domain of cats
because it has been trained on other images, such as dogs).

6
 What is a machine
learning (ML) model?
Example: Spam detection for email
Input Deployed ML model Output
new email spam detection valid email or spam

valid

spam

Labeled historical data Machine learning algorithm

valid and spam emails one of multiple possible
candidate algorithms

Machine learning model training

iterative process of experimenting and testing

Although a significant amount of time is spent obtaining and preparing data for machine learning (ML), the bulk of
the ML workflow is focused on building, training, deploying, and monitoring “models.” The term “models” gets
used frequently and is a core concept.

The essence of machine learning can be boiled down to three things:

1. An organization has historical data

2. Patterns exist in this data
3. Once a pattern is identified, it can be used to make predictions about future actions or behaviors

For people familiar with Yahoo or Google Mail, they will have probably seen a folder named “Junk” or “Spam.”
When looking inside this folder, people generally find it contains email messages they didn’t really want to receive.
When people continuously throw away messages that come from the same sender without reading them, at some
point they will discover that these messages are automatically placed in the Junk/Spam folder by their email
provider. It’s a good bet that an ML algorithm was used create a model that performs these “spam filtering” tasks.
Therefore, spam filtering is an excellent example to illustrate what a machine learning model is.

In the illustration shown on this slide, a learning model consists of two components – an algorithm a data scientist
chose in order to find a pattern that exists (in this case, a pattern that indicates whether an incoming email
message is spam), based on the probability that it will do the task it’s expected to do once it has been trained.

ML models can be thought of as having two phases: a training phase and a prediction or scoring phase. Going back
to the email spam filtering example, a typical spam email message contains words like “sweepstakes,”
“inheritance,” or “Viagra.” On the other hand, these words probably don’t appear very often in emails people are
likely to be considered valid. During the model training phase, the data scientist would tell the chosen algorithm to

7
look at historical data and note the probability with which each of these words appear in both
spam and valid email messages.

For example, if it is known that 6,000 out of 10,000 messages in the training data set are
spam and if the word “sweepstakes” appears in 2,000 spam emails and 10 valid emails, the
probability that a spam email message contains the word sweepstakes is 0.333 (2000/6000 =
0.333). And the probability that this word will appear in a valid email message is 0.0025
(10/4000 = 0.0025). Thus, if an incoming email message contains the word “sweepstakes,”
the probability that the message is spam is much higher than the probability that the message
is valid.

Notes:

• Machine learning (ML) refers to a broad set of techniques to train a computer to learn from
its inputs, using existing data, and one or more “training” methods, instead of being
explicitly programmed. ML helps a computer to achieve AI.

• An algorithm is a procedure used for solving a problem or performing a computation.

Algorithms act as an exact list of instructions that conduct a sequence of specified actions
in either hardware- or software-based routines.

7
 Machine learning (ML) types
Classification models
Example: Spam detection for email

Input Deployed ML model Output

Classification models assign labels to model new email spam detection valid email or spam
inputs or assign them to specific categories.
Common use cases include: valid

• Fraud detection: predict whether a transaction

is fraudulent based on patterns in the data spam

• Sentiment analysis: classify text as

positive, negative, or neutral
• Medical diagnosis: assign a disease
label to a patient’s case, based on
symptoms and medical history
• Image recognition: recognize objects
Labeled historical data Machine learning algorithm
or identify people based on visible valid and spam emails
features and characteristics
Machine learning model training
iterative process of experimenting and testing

Classification models are widely used in various fields to predict a specific label or category based on one or more
input features.

These are some common use cases for classification models:

• Spam detection: Classification models can be used to classify emails as spam or not spam based on their
content, sender, and other factors.

• Fraud detection: Classification models can be used to detect fraudulent transactions, insurance claims, or
other activities based on patterns in data.

• Sentiment analysis: Classification models can be used to classify text as positive, negative, or neutral
sentiment based on the words used and the context.

• Medical diagnosis: Classification models can be used to diagnose diseases based on symptoms, medical
history, and other factors.

• Image recognition: Classification models can be used to recognize objects or identify people in images or
videos based on their features and characteristics.

• Credit scoring: Classification models can be used to predict the likelihood of default on a loan or credit card
based on the applicant's credit history, income, and other factors.

• Customer segmentation: Classification models can be used to segment customers into different groups based
on their behavior, preferences, and demographics.

8
In summary, classification models are a powerful tool for predicting discrete labels or
categories based on input features and have a wide range of applications in fields such as
finance, healthcare, marketing, and more.

Note:

• Machine learning (ML) refers to a broad set of techniques to train a computer to learn from
its inputs, using existing data, and one or more “training” methods, instead of being
explicitly programmed. ML helps a computer to achieve AI.

8
 Machine learning (ML) types
Regression models
Example: House price prediction

Input Deployed ML model Output

Regression models make predictions address & date house price predictor house price
based on the model input.
Common use cases include:
• Stock market analysis: securities
price prediction based on historical
data or news events
• Sales: forecasting based on
historical data or market trends
• Healthcare: predict patient outcomes
based on factors such as age, gender,
medical history, or treatment plans
Labeled historical data Machine learning algorithm
• Customer behavior analysis: predict real estate market data
future customer purchasing patterns
based on demographic data, past purchase
Machine learning model training
history, and advertising campaigns iterative process of experimenting and testing

In machine learning (ML), regression refers to a type of supervised learning algorithm that is used to predict a
numerical value (the target) based on a set of input features. The example on this slide is a regression model that
predicts house prices.

Regression models are trained with labeled (the supervised part) historical data, for example, real estate market
data where the target field, which is the sale sale price of houses is listed alongside multiple other “features” (data
fields of interest), like square footage, number of bedrooms and bathrooms, address, sale date, and number of
days on market.

The goal of a regression model is to identify the relationship between the input features and the target variables. In
this case, when a house price prediction model is being trained, it looks for relationships between the various
features listed earlier (like address) with the target value (house price). Once it’s been trained, the model then
uses this relationship to predict the house price for new input data (now houses the model will evaluate that it’s
never seen before).

Regression models are widely used in many fields, including finance, healthcare, and engineering, where they can
be used to predict values such as stock prices, patient outcomes, sales forecasts, or customer behavior based on
input features.

Notes:

• Machine learning (ML) refers to a broad set of techniques to train a computer to learn from its inputs, using
existing data, and one or more “training” methods, instead of being explicitly programmed. ML helps a
computer to achieve AI.

9
• An algorithm is a procedure used for solving a problem or performing a computation.
Algorithms act as an exact list of instructions that conduct a sequence of specified actions
in either hardware- or software-based routines.

9
 Machine learning (ML) types
Clustering models
Example: Market basket analysis

Clustering models identifies distinct groupings

of individual data points that share common
characteristics within a larger data set.
Common use cases include:
• Customer segmentation: group customers
Unlabeled Clustering Output
based on similar preference, behaviors, historical data algorithm items commonly
and demographics transaction data purchased together
• Genetic analysis: group genes with
similar functions or processes
• Social network analysis: identify communities
or groups within a social network
• Market basket analysis: identify items
that are commonly purchased together

Clustering is a type of unsupervised learning where data is grouped into clusters based on similarities between
data points.

These are some common use cases for clustering:

• Customer segmentation: Clustering can be used to group customers with similar behaviors, preferences, and
demographics for targeted marketing campaigns and personalized recommendations.

• Image segmentation: Clustering can be used to segment images into different regions based on color, texture,
and other features.

• Anomaly detection: Clustering can be used to identify outliers or anomalies in data that do not belong to any of
the existing clusters.

• Document clustering: Clustering can be used to group similar documents together for better organization and
searchability.

• Genetic analysis: Clustering can be used to group genes with similar functions or characteristics for better
understanding of genetic processes.

• Social network analysis: Clustering can be used to identify communities or groups within a social network
based on connections between individuals.

• Market basket analysis: Clustering can be used to identify groups of items that are often purchased together in
a market basket analysis.

10
In summary, clustering is a powerful tool for unsupervised learning and has a wide range of
applications in fields such as marketing, biology, social sciences, and more.

Note:

• Machine learning (ML) refers to a broad set of techniques to train a computer to learn from
its inputs, using existing data, and one or more “training” methods, instead of being
explicitly programmed. ML helps a computer to achieve AI.

10
 Machine learning (ML) types
Generative models
Example: Text generation

Generative models create new data in Input Deployed Output

response to input requests (prompts). prompt foundation model answer

Common use cases include:

• Text generation: generate new text
in response to a prompt, summarizing
text, or writing a lengthy essay
• Code generation: generate computer
code based on a textual description
of the proposed program
• Image generation: create images
based on the prompt
Unlabeled text data Machine learning algorithm

Machine learning model training

iterative process of experimenting and testing

A relatively new kind of machine learning (ML) model is a generative model. In this case, rather than having a
specific kind of output (like a prediction or classification), the output can be far further ranging. For example, the
most common kinds of generative models are for text, which can be designed for human-readable text, computer
code, or synthetically generated data sets. There are also generative models, which can generate binary data, like
images, video, and music, all based on an input prompt which asks the AI for something.

The pattern for how generative models are trained is similar to other kinds of ML models: in the case of a
foundation model, very large sets of unlabeled text data train the model, which is then deployed. Text-based
foundation models (which are specifically known as large language models (LLMs), like the widely used ChatGPT)
are frequently deployed in chatbot applications, where people can submit questions to them, and receive answers.

Notes:

• Machine learning (ML) refers to a broad set of techniques to train a computer to learn from its inputs, using
existing data, and one or more “training” methods, instead of being explicitly programmed. ML helps a
computer to achieve AI.

• Generative AI refers to a set of AI algorithms that can generate new outputs — such as text, images, code, or
audio — based on the training data, unlike traditional AI systems that are designed to recognize patterns and
make predictions. Sometimes the AI that powers these solutions are referred to as decoders.

• An algorithm is a procedure used for solving a problem or performing a computation. Algorithms act as an
exact list of instructions that conduct a sequence of specified actions in either hardware- or software-based
routines.

11
• Foundation models are typically built using a specific kind of neural network architecture,
called a transformer, which is designed to generate sequences of related data elements (for
example, like a sentence).

• A transformer model is a neural network architecture useful for understanding language,

which does not have to understand words one at a time but can look at an entire sentence
at once for context and disambiguation. A technique called self-attention allows these
models to focus on the particular words that are important in understanding the meaning of
the sentence. Transformers single-handedly spawned the world of foundation models and
its birthplace research paper (Attention is all You Need) spawned the AI inflection point
that is happening today. AI transformer models can be Encoder models (like BERT),
Decoder models (like GPT), or Encoder and Decoder models (like T5). Learn more about
Transformers at: https://www.youtube.com/watch?v=ZXiruGOCn9s.

• Neural networks are a set of algorithms, modeled loosely after the neural networks found
in the human brain, that are designed to recognize hidden patterns in data.

• A large language model (LLM) is a type of machine learning model that has been trained
on large quantities of unlabeled text using self-supervised learning and can perform a
variety of natural language processing (NLP) tasks (even when that language is a
programming language). Output may range from books, articles, social media posts, online
conversations, and even code. The architecture of an LLM consists of layers of neural
networks that learn to generate language in a way that is similar to how humans use
language

• ChatGPT is an AI-generative engine that automatically generates text based on written

prompts in a fashion that is very advanced, creative, and conversational in nature. AI
research lab OpenAI launched it for public use on November 30, 2022. GPT stands for
Generative Pre-trained Transformer and refers to a family of neural network-based large
language models (LLMs) developed by OpenAI.

• Natural language processing (NLP) is the technology that gives computers the ability to
understand text and spoken words in much the same way human beings can. NLP
combines computational linguistics — rule-based modeling of human language — with
statistical, machine learning (ML), and deep learning (DL) models. These technologies
enable computers to process human language in the form of text or voice data and to
‘understand’ its full meaning, complete with the speaker or writer’s intent and sentiment.

• Deep learning (DL) is a technique for implementing machine learning (ML) that relies on
deep artificial neural networks to perform complex tasks such as image recognition, object
detection, and natural language processing (NLP). Neural networks are a set of algorithms,
modeled loosely after the neural networks found in the human brain, which are designed to
recognize hidden patterns in data. The “deep” in DL refers to a neural network comprised of
more than three layers (which include the input and the output layers).

11
 Machine learning
operations (MLOps)
lifecycle Monitor model Find data

Use math, statistics, specialized

programming, and analytics to
extract meaningful insights
from data.
MLOps lifecycle
Designed to build, deploy, Deploy model Prepare data
manage, and monitor
ML models.

Test model Train model

Companies build, train, deploy, and use machine learning (ML) and deep learning (DL) models by going through the
steps illustrated on this slide. This is known as the machine learning operations (MLOps) workflow.

There is a great deal more to developing artificial intelligence (AI) models than ML, and it involves people from
different areas such as business-people, data engineers, data stewards, data scientist, MLOps people, and others.
This is a team effort that requires smooth communication and information sharing.

One key activity is finding the data required to create the machine learning model. This is done with the help of
data stewards and data engineers. The data could be from within the enterprise or come from public sites such as
federal, state, or city governments. The data needs to be examined and cleaned up. The resulting understanding
can uncover additional needs that would add additional requirements. Then the data sources and data needs to be
catalogued and privacy rules need to be applied so only the appropriate people will see the exact data.

The data processing is not over. Some features may need to be converted from character to numerical values. For
example, if the business must predict the number of widgets that will be sold next month, it might be required to
specify how many of each available color. So, the colors red, white, and blue are converted to numbers with a
special processing indicating these are categorical values. If there is a need to predict how many resources are
needed on a future Tuesday, it might be required to extract the day of the week from the training records dates.
These type of conversions are called feature engineering.

The iterative process of data collection, understanding, and preparation consumes around 80% of the time spent
in data science projects. Often, the first look at the enterprise data results in a rude awakening where the
enterprise data may not have the quality expected.

After the proper data with the desired feature engineering is collected and preforms, it is time to create the model

12
where the data scientist can try multiple algorithms with different hyper parameters. The
resulting models are evaluated, and the best model chosen for deployment. The MLOps
people get involved in the model deployment and monitoring so feedback can be returned,
and a decision can be made as to when the model needs to be refreshed or retired altogether.

In brief, many people’s expertise and experience is applied in the process to go from a
business problem to a successfully deployed machine learning model.

Notes:

• Machine learning (ML) refers to a broad set of techniques to train a computer to learn from
its inputs, using existing data, and one or more “training” methods, instead of being
explicitly programmed. ML helps a computer to achieve AI.

• Deep learning (DL) is a technique for implementing machine learning (ML) that relies on
deep artificial neural networks to perform complex tasks such as image recognition, object
detection, and natural language processing (NLP). Neural networks are a set of algorithms,
modeled loosely after the neural networks found in the human brain, which are designed to
recognize hidden patterns in data. The “deep” in DL refers to a neural network comprised of
more than three layers (which include the input and the output layers).

• Natural language processing (NLP) is the technology that gives computers the ability to
understand text and spoken words in much the same way human beings can. NLP
combines computational linguistics — rule-based modeling of human language — with
statistical, machine learning (ML), and deep learning (DL) models. These technologies
enable computers to process human language in the form of text or voice data and to
‘understand’ its full meaning, complete with the speaker or writer’s intent and sentiment.

• Neural networks are a set of algorithms, modeled loosely after the neural networks found
in the human brain, that are designed to recognize hidden patterns in data.

• An algorithm is a procedure used for solving a problem or performing a computation.

Algorithms act as an exact list of instructions that conduct a sequence of specified actions
in either hardware- or software-based routines.

• Parameters are numerical values that define a large language model’s (LLM) structure and
behavior, like clues that help it guess what words come next. Modern systems like ChatGPT
are thought to have hundreds of billions of parameters.

12
 Machine learning Supervised Semi-supervised
learning learning
(ML) methods
An operator provides By using a combination
the ML algorithm with of labeled and
a known dataset that unlabeled data, ML
includes desired algorithms can learn to
inputs and outputs. label unlabeled data.

Unsupervised Reinforcement
learning learning

The machine determines A technique that

the correlations and teaches an AI model
relationships by to find the best result
analyzing available data. by trial and error.

Machine learning (ML) uses a broad set of techniques to train a computer to learn from its inputs, using existing
data, and one or more “training” methods, instead of being explicitly programmed. ML helps a computer to achieve
artificial intelligence (AI), gradually improving its accuracy over time.

ML helps computer systems learn and improve from experience. As more and more data is inputted into a
machine, it helps the algorithms teach the computer, thus improving the delivered results. For example, if a user
asks Amazon’s Alexa to play their favorite music station on their Amazon Echo device, Alexa will go to the station
the user has played most often — something Alexa inferred by observing all that times that user listened to music,
what kind of music they listened, to, when they listened to music, for how long, what where they doing while
listening to music (working our or driving), and of course what music streaming channel they listened to the most.
A user can further improve and refine their listening experience by telling Alexa to skip songs, adjust the volume,
and other commands (which means Alexa has been trained to understand the tasks of music listening).

There are four main types of ML algorithms: supervised, semi-supervised, unsupervised, and reinforcement
learning.

In supervised learning, an operator provides the ML algorithm with a known dataset that includes desired inputs
and outputs (referred to as labelled data). The algorithm identifies patterns in data, learns from observations, and
makes predictions. The operator corrects the algorithm, and the process continues until the algorithm achieves a
high level of accuracy. Supervised learning is very accurate but can be very very expensive because a lot of time
must be spent labelling the data outputs to inputs; when taking into account the sheer volume of data needed to
train AI, this has proven onerous and as a source of friction for many companies on their AI journey because of the
time and cost to label data.

Under the umbrella of supervised learning fall: classification, regression, and forecasting.

13
• In classification tasks, the ML program must draw a conclusion from observed values and
determine to what category new observations belong. For example, when filtering emails
as ‘spam’ or ‘not spam,’ the program must look at existing observational data and filter the
emails accordingly.

• In regression tasks, the ML program must estimate and understand the relationships
among variables. Regression analysis focuses on one dependent variable and a series of
other changing variables, making it particularly useful for prediction and forecasting. For
example, how does the number of bedrooms, proximity to schools (and so on) affect the
price of a house for sale.

• Forecasting is the process of making predictions about the future based on the past and
present data and is commonly used to analyze trends. For example, at certain
temperatures, people are more likely to buy certain products and restaurants with
dynamic menus will change their drive-thru menus based on these predictions to
maximize sales.

Semi-supervised learning uses both labeled and unlabeled data. Labeled data is essentially
information that has meaningful tags so that the algorithm can understand the data, while
unlabeled data lacks that information. By using this combination, ML algorithms can learn to
label unlabeled data.

In unsupervised learning, the ML algorithm studies data to identify patterns. There is no

answer key or human operator to provide instruction. Instead, the machine determines the
correlations and relationships by analyzing available data. In this process, the ML algorithm is
left to interpret large data sets and address that data accordingly. The algorithm tries to
organize that data in some way to describe its structure. This might mean grouping the data
into clusters or arranging it in a way that looks more organized. As it assesses more data, the
algorithm’s ability to make decisions on that data gradually improves and becomes more
refined. Under the umbrella of unsupervised learning, fall: clustering and dimension reduction.

• Clustering involves grouping sets of similar data (based on defined criteria). It’s useful for
segmenting data into several groups and performing analysis on each data set to find
patterns. For example, an insurance algorithm that looks at a health claim and
automatically adjudicates it for payment… perhaps the claim is rejected, and the AI moves
it to a follow-up process, or flags it for fraud.

• Dimension reduction reduces the number of variables (features) being considered to find
the exact information required. This is a key component because the more dimensions
used, the more complex an algorithm becomes (and potentially the more prone to over-
fitting it becomes too).

Reinforcement learning is a technique that teaches an AI model to find the best result by trial
and error. A ML algorithm is provided with a set of actions, parameters, and end values. By

13
defining the rules, the ML algorithm then tries to explore different options and possibilities,
monitoring and evaluating each result to determine which one is optimal. Reinforcement
learning teaches the machine by trial and error, receiving rewards (or punishment) from an
algorithm based on its results. It learns from past experiences and begins to adapt its
approach in response to the situation to achieve the best possible result. These systems can
be enhanced by humans giving feedback on its performance. A great example of
Reinforcement learning is watching the space of game playing by computers (from board
games like Chess and Go to video games like Super Mario Brothers or Atari Breakout, among
others).

Notes:

• An algorithm is a procedure used for solving a problem or performing a computation.

Algorithms act as an exact list of instructions that conduct a sequence of specified actions
in either hardware- or software-based routines.

• Parameters are numerical values that define a large language model’s (LLM) structure and
behavior, like clues that help it guess what words come next. Modern systems like ChatGPT
are thought to have hundreds of billions of parameters.

• ChatGPT is an AI-generative engine that automatically generates text based on written

prompts in a fashion that is very advanced, creative, and conversational in nature. AI
research lab OpenAI launched it for public use on November 30, 2022. GPT stands for
Generative Pre-trained Transformer and refers to a family of neural network-based large
language models (LLMs) developed by OpenAI.

• Amazon’s Alexa is a virtual assistant technology. The cloud-based voice service is available
on more than 100 million devices from Amazon and third-party device manufacturers.

Sources:

• What is machine learning? Webpage: https://www.ibm.com/topics/machine-learning

• What is unsupervised learning? Webpage: https://www.ibm.com/topics/unsupervised-

learning

• A guide to the types of machine learning algorithms and their applications Webpage:
https://www.sas.com/en_gb/insights/articles/analytics/machine-learning-algorithms.html

13
 Large language
models (LLMs)
LLMs are machine learning
models that have been
trained on large quantities
of unlabeled text using
self-supervised learning
and can perform a variety
of natural language
processing tasks .

ChatGPT is an example of
a generative AI chatbot
developed using an LLM.

Practically overnight, OpenAI’s ChatGPT released the powers of artificial intelligence (AI) and large language
models (LLMs) from the abstract into something more graspable. ChatGPT is a generative AI chatbot developed by
the OpenAI research lab. GPT stands for Generative Pre-trained Transformer (GPT) and refers to a family of neural
network-based LLMs developed by OpenAI. It was trained on a large body of text from a variety of sources
(Wikipedia, 3-star or above ratings on Reddit reviews, and more). Once it answers a question, the individual asking
that question can amend, reword, reorganize, probe deeper, or even rewrite it in a different language style by
interacting with this AI agent.

An LLM is a type of AI model that has been trained on large quantities of unlabeled text using self-supervised
learning and can perform a variety of natural language processing (NLP) tasks. The text data may range from
books, articles, social media posts, and online conversations (but LLMs can also learn the language of code). The
architecture of an LLM consists of layers of neural networks that learn to generate language in a way that is similar
to how humans use language. NLP tasks include generating and classifying text, answering questions in a
conversational manner, and translating text from one language to another.

Parameters are numerical values that define an LLM’s structure and behavior, like clues that help it guess what
words come next. Modern systems like ChatGPT have hundreds of billions of parameters. These parameters are in
place to train the system. Although, bias and hallucinations are still a concern for these LLMs.

Notes:

• Generative AI refers to a set of AI algorithms that can generate new outputs — such as text, images, code, or
audio — based on the training data, unlike traditional AI systems that are designed to recognize patterns and
make predictions. Sometimes the AI that powers these solutions are referred to as decoders.

14
• GPT stands for Generative Pre-trained Transformer and refers to a family of neural
network-based large language models (LLMs) developed by OpenAI; for example, GPT-3
and GPT-4.

• Natural language processing (NLP) is the technology that gives computers the ability to
understand text and spoken words in much the same way human beings can. NLP
combines computational linguistics — rule-based modeling of human language — with
statistical, machine learning (ML), and deep learning (DL) models. These technologies
enable computers to process human language in the form of text or voice data and to
‘understand’ its full meaning, complete with the speaker or writer’s intent and sentiment.

• Machine learning (ML) refers to a broad set of techniques to train a computer to learn from
its inputs, using existing data, and one or more “training” methods, instead of being
explicitly programmed. ML helps a computer to achieve AI.

• Deep learning (DL) is a technique for implementing machine learning (ML) that relies on
deep artificial neural networks to perform complex tasks such as image recognition, object
detection, and natural language processing (NLP). Neural networks are a set of algorithms,
modeled loosely after the neural networks found in the human brain, which are designed to
recognize hidden patterns in data. The “deep” in DL refers to a neural network comprised of
more than three layers (which include the input and the output layers).

• A transformer model is a neural network architecture useful for understanding language,

which does not have to understand words one at a time but can look at an entire sentence
at once. A technique called self-attention allows the model to focus on the particular words
that are important in understanding the meaning of the sentence. Transformers single-
handedly spawned the world of foundation models and its birthplace research paper
(Attention is all You) spawned the AI inflection point that is happening today. AI
transformer models can be Encoder models (like BERT), Decoder models (like GPT), or
Encoder and Decoder models (like T5). Learn more about Transformers at:
https://www.youtube.com/watch?v=ZXiruGOCn9s.

• Foundation models are typically built using a specific kind of neural network architecture,
called a transformer, which is designed to generate sequences of related data elements (for
example, like a sentence).

• Neural networks are a set of algorithms, modeled loosely after the neural networks found
in the human brain, that are designed to recognize hidden patterns in data.

• An algorithm is a procedure used for solving a problem or performing a computation.

Algorithms act as an exact list of instructions that conduct a sequence of specified actions
in either hardware- or software-based routines.

• Bias is a type of error that can occur if an artificial intelligence (AI) models’ output is
skewed by the model’s training data. For example, a model may associate specific traits or
professions with a certain race or gender, leading to inaccurate predictions and offensive

14
 responses.

• Hallucination is a well-known phenomenon in large language models (LLMs) in which the

system provides an answer that is factually incorrect, irrelevant, or nonsensical because of
limitations in its training data and architecture; more concerning is the hallucinated answer
sounds plausible.

14
 What’s next with foundation models:

Conventional AI Zero-shot prompting Few-shot prompting Data-driven tuning

1000s – 100000s labeled No labeled data needed 1-10 labeled data points / 100s – 1000s labeled
data points / task task data points / task

Task Task Task Task Task Task

Task Task

Labeled tuning data

Model Model Handcrafted examples

Labeled Labeled
training training
Pre-trained generative foundation model
data data

Conventional artificial intelligence (AI) techniques involve task-specific models being trained with large sets of
labeled training data. A recent innovation, known as foundation models, is creating a massive disruption in this
conventional AI approach.

Foundation models are large general-purpose models that are trained on extremely large unlabeled data sets,
using transformer models. These models can be customized for down-stream tasks (fine-tuned) with small sets of
labeled data and yield exceptional customization results for tasks that a trained foundation model wasn’t trained
on.

And while the underlying model is generative, these models can be adapted to perform any number tasks (referred
to as down-stream tasks), including those not normally thought of as being generative.

What is Prompting?
Foundation models can be “instructed” to solve various natural language processing (NLP) tasks through a process
called “prompt engineering” (some people will refer to this as instruction promoting). Prompt engineering is a
process of designing, creating, and testing prompts for natural language generation systems. Specifically, prompt
engineering (also known as In-Context Prompting), refers to methods for how to communicate with large language
models (LLMs) to steer its behavior for desired outcomes without updating the model weights. At its core, the goal
of prompt engineering is about alignment and model steerability.

Just what is a prompt? A prompt is a piece of text inserted as input examples to a LLM so that the original task can
be formulated as a (masked) language modeling problem. With the right mix of creativity and technological skill,
these powerful tools can be combined to create text generation models with impressive accuracy and flexibility for
all types of applications. In fact, the Prompt Engineer is a super-hot new job that has been created because of
generative AI. Prompts are the inputs into the algorithm and especially because of their pricing mechanics (clients
often get charged for how much gets built, or for the frequency of submissions) it’s important to get the right

15
combinations of prompts to get the full value out of a cycle of the model. Truthfully, prompt
engineering is somewhat of an “art”: it requires experimentation to achieve good results and is
an active area of research and discussion. That said, hand-crafting good prompts can be tricky,
requiring domain expertise, and the outcome can be unintuitive.

There are three types of prompting:

1. Few-shot prompting: Few-shot learning presents a set of high-quality demonstrations,

each consisting of both input and desired output, on the target task. In this form of
prompting, the model receives a few (two, or up to a few tens) examples of input text
along with the desired output… the model is then prompted to generate the output for
the unseen text. The prompt must be “engineered” appropriately for the specific task.
Few-shot prompting is a technique where the model is given a small number of
examples, typically between two and five, in order to quickly adapt to new examples of
previously seen objects.

Few-shot learning can be used, in the context of prompt engineering, to create natural
language text with a limited amount of input data. Although it requires less data, this
technique can allow for the creation of more versatile and adaptive text generation
models. By using advanced techniques such as few-shot prompting, it is possible to
create natural language generation models that are more flexible, adaptable, and
engaging for human users. For example, a prompt engineer might give the LLM three
examples of the output structure they want back from the model; then the model will
guess with much higher accuracy and reliability of what is wanted and return text based
on the given examples and its training about text.

As the model first sees good examples, it can better understand the prompt’s intention
and criteria for what kinds of answers are wanted. Therefore, few-shot learning often
leads to better performance than zero-shot. However, it comes at the cost of more token
consumption and may hit the context length limit for some LLMs when input and output
text are long.

2. One shot prompting: In this prompting method, a prompt engineer gives the LLM just a
single example of the output structure they would like to get back. The model will guess
what is wanted and return text on the example and its training about text.

3. Zero shot prompting : the model is provided with a description of the desired output, but
no examples! Basically, the prompt engineer just gives the LLM a task to complete
without any instructions. The model then guesses what the prompt engineer wants and
returns some text (or translation in this slide’s example) based on its training and
understanding about text.

Intuitively, in “few-shot” mode, a prompt engineer has more control over the output by
steering the model towards the desired output. Consider this example:

15
• Input : “Today U.S. President Joseph R. Biden, Jr. and IBM (NYSE: IBM) Chairman and CEO
Arvind Krishna will tour IBM's Poughkeepsie New York site to see firsthand where the
future of computing is being innovated, designed and manufactured.”

• Input: IBM: company, Joseph R. Biden, Jr.: person, Arvind Krishna: person, Armonk, US:
location, Poughkeepsie: location, New York: location

• Prompt: Give me the entities from the following: Microsoft has seen declining market share
in the EU, Eric Schmidt said.

• Model Response: EU: country, Eric Schmidt: person

Note how the input to the LLM included examples of entitles (labelled data) and the LLM was
able to understand and draw conclusions that if IBM is a company than Microsoft must be a
company too; if US is a country/location, then the European Union (EU) must be one too; and if
Arvind Krishna is a person, it’s likely that Eric Schmidt is one too.

Learn more about prompting at: https://www.allabtai.com/prompt-engineering-tips-zero-one-

and-few-shot-prompting/ or https://thegradient.pub/prompting/.

Data-driven tuning:
Tuning is the process of maximizing a model's performance without overfitting or creating too
high of a variance. In machine learning (ML), this is accomplished by selecting appropriate
hyperparameters. Hyperparameters can be thought of as the “dials” or “knobs” of a ML model
that are used to set up the tuning process. The data-driven approach establishes, first and
foremost, which key performance indicators (KPIs) are relevant to the company's business. It
creates a group of decision-makers to make that data intelligible and thus truly exploitable.

Notes:

• Foundation models are typically built using a specific kind of neural network architecture,
called a transformer, which is designed to generate sequences of related data elements (for
example, like a sentence).

• Neural networks are a set of algorithms, modeled loosely after the neural networks found
in the human brain, that are designed to recognize hidden patterns in data.

• A transformer model is a neural network architecture useful for understanding language,

which does not have to understand words one at a time but can look at an entire sentence
at once. A technique called self-attention allows the model to focus on the particular words
that are important in understanding the meaning of the sentence. Transformers single-
handedly spawned the world of foundation models and its birthplace research paper
(Attention is all You) spawned the AI inflection point that is happening today. AI
transformer models can be Encoder models (like BERT), Decoder models (like GPT), or
Encoder and Decoder models (like T5). Learn more about Transformers at:
https://www.youtube.com/watch?v=ZXiruGOCn9s.

• An algorithm is a procedure used for solving a problem or performing a computation.

15
 Algorithms act as an exact list of instructions that conduct a sequence of specified actions
in either hardware- or software-based routines.

• Natural language processing (NLP) is the technology that gives computers the ability to
understand text and spoken words in much the same way human beings can. NLP
combines computational linguistics — rule-based modeling of human language — with
statistical, machine learning (ML), and deep learning (DL) models. These technologies
enable computers to process human language in the form of text or voice data and to
‘understand’ its full meaning, complete with the speaker or writer’s intent and sentiment.

• A large language model (LLM) is a type of machine learning model that has been trained
on large quantities of unlabeled text using self-supervised learning and can perform a
variety of natural language processing (NLP) tasks (even when that language is a
programming language). Output may range from books, articles, social media posts, and
online conversations. The architecture of an LLM consists of layers of neural networks that
learn to generate language in a way that is similar to how humans use language.

• Natural language generation (NLG) is a subtopic of natural language processing (NLP) that
is the process of generating text that appears to be written by a human, without the need
for a human to actually write it.

• Generative AI refers to a set of AI algorithms that can generate new outputs — such as
text, images, code, or audio — based on the training data, unlike traditional AI systems that
are designed to recognize patterns and make predictions. Sometimes the AI that powers
these solutions are referred to as decoders.

• Machine learning (ML) refers to a broad set of techniques to train a computer to learn from
its inputs, using existing data, and one or more “training” methods, instead of being
explicitly programmed. ML helps a computer to achieve AI.

• Deep learning (DL) is a technique for implementing machine learning (ML) that relies on
deep artificial neural networks to perform complex tasks such as image recognition, object
detection, and natural language processing (NLP). Neural networks are a set of algorithms,
modeled loosely after the neural networks found in the human brain, which are designed to
recognize hidden patterns in data. The “deep” in DL refers to a neural network comprised of
more than three layers (which include the input and the output layers).

• Parameters are numerical values that define a large language model’s (LLM) structure and
behavior, like clues that help it guess what words come next. Modern systems like ChatGPT
are thought to have hundreds of billions of parameters.

15
 Traditional AI
applications
Speech Customer
recognition service

Computer Recommendation Automated stock

vision engines trading

Artificial intelligence (AI) is changing the way businesses operate in fundamental ways, from how they
communicate with their customers through virtual assistants, to automating key workflows, and even managing
network security. Companies around the world have accelerated their rollout of AI and businesses plan to invest in
all areas of AI, from skills and workforce development to buying AI tools and embedding those into their business
processes. There are numerous, real-world applications of AI systems today. These are some of the most common
examples:

• Speech recognition: This technology, also known as automatic speech recognition (ASR), computer speech
recognition, or speech-to-text, is a capability that uses natural language processing (NLP) to process human
speech into a written format. NLP is one of the foundational technologies for AI models and has steadily
become one of the most important and commonplace tools for organizations to communicate with customers
and empower their employees. Many mobile devices incorporate speech recognition into their systems to
conduct voice search or allow more accessibility for texting, such as Apple’s Siri.

• Customer service: Online virtual agents are replacing (or augmenting) human agents along the customer
journey. They answer frequently asked questions (FAQs) around topics, like shipping, or provide personalized
advice, cross-selling products, or suggesting sizes for users, changing customer engagement across websites
and social media platforms. Examples include messaging bots on e-commerce sites with virtual agents,
messaging applications, such as Slack and Facebook Messenger, and tasks usually done by virtual assistants
and voice assistants. There are also invisible workflows like suspending a lost credit card.

• Computer vision: This AI enables computers and systems to derive meaningful information from digital
images, videos, and other visual inputs, and based on those inputs, it acts. Powered by convolutional neural
networks (CNN), computer vision has applications within photo tagging in social media, radiology imaging in
healthcare, self-driving cars within the automotive industry, and more. For example, Maximo Visual Inspector
(MVI) lets non-technical business users build power CNNs for tasks such as inspection — like vegetation growth

16
 for utilities endangering powerlines or weld joints on a manufacturing car line.

• Recommendation engines: Using past consumption behavior data, AI algorithms can help
to discover data trends that can be used to develop more effective cross-selling strategies.
This AI application is used to make relevant add-on recommendations to customers during
the checkout process for online retailers. This technology is ubiquitous with video
streamers such as Netflix, but the approach is used everywhere in lesser know industries
too; that said, no matter they industry, the approach is the same.

• Automated stock trading: Designed to optimize stock portfolios, AI-driven, high-frequency

trading platforms make thousands or even millions of trades per day without human
intervention.

Notes:

• Natural language processing (NLP) is the technology that gives computers the ability to
understand text and spoken words in much the same way human beings can. NLP
combines computational linguistics — rule-based modeling of human language — with
statistical, machine learning (ML), and deep learning (DL) models. These technologies
enable computers to process human language in the form of text or voice data and to
‘understand’ its full meaning, complete with the speaker or writer’s intent and sentiment.

• Neural networks are a set of algorithms, modeled loosely after the neural networks found
in the human brain, that are designed to recognize hidden patterns in data.

• An algorithm is a procedure used for solving a problem or performing a computation.

Algorithms act as an exact list of instructions that conduct a sequence of specified actions
in either hardware- or software-based routines.

• Deep learning (DL) is a technique for implementing machine learning (ML) that relies on
deep artificial neural networks to perform complex tasks such as image recognition, object
detection, and natural language processing (NLP). Neural networks are a set of algorithms,
modeled loosely after the neural networks found in the human brain, which are designed to
recognize hidden patterns in data. The “deep” in DL refers to a neural network comprised of
more than three layers (which include the input and the output layers).

• Machine learning (ML) refers to a broad set of techniques to train a computer to learn from
its inputs, using existing data, and one or more “training” methods, instead of being
explicitly programmed. ML helps a computer to achieve AI.

Source:

• What is artificial intelligence? Webpage: https://www.ibm.com/topics/artificial-intelligence

16
 Natural
language
processing

The technology that

gives computers the
ability to understand
text and spoken words
in much the same
way humans can.

Natural language processing (NLP) is the technology that gives computers the ability to understand text and
spoken words in much the same way humans can. NLP combines computational linguistics—rule-based modeling
of human language—with statistical, machine learning (ML), and deep learning (DL) models. These technologies
enable computers to process human language in the form of text or voice data and to make sense of its full
meaning, complete with the speaker or writer’s intent and sentiment.

NLP drives computer programs that translate text from one language to another, respond to spoken commands,
and summarize large volumes of text rapidly—even in real time. NLP technology is used in voice-operated global
positioning system (GPS) systems, digital assistants, speech-to-text dictation software, customer service chatbots,
and other consumer conveniences. But NLP also plays a growing role in enterprise solutions that help streamline
business operations, increase employee productivity, and simplify mission-critical business processes.

NLP is usually used as a top category for all AI-based language processing like natural language understanding
(NLU) and natural language generation (NLG).

Notes:

• Machine learning (ML) refers to a broad set of techniques to train a computer to learn from its inputs, using
existing data, and one or more “training” methods, instead of being explicitly programmed. ML helps a
computer to achieve AI.

• Deep learning (DL) is a technique for implementing machine learning (ML) that relies on deep artificial neural
networks to perform complex tasks such as image recognition, object detection, and natural language
processing (NLP). Neural networks are a set of algorithms, modeled loosely after the neural networks found in
the human brain, which are designed to recognize hidden patterns in data. The “deep” in DL refers to a neural

17
 network comprised of more than three layers (which include the input and the output
layers).

• Neural networks are a set of algorithms, modeled loosely after the neural networks found
in the human brain, that are designed to recognize hidden patterns in data.

• An algorithm is a procedure used for solving a problem or performing a computation.

Algorithms act as an exact list of instructions that conduct a sequence of specified actions
in either hardware- or software-based routines.

• Natural language understanding (NLU) is a subtopic of natural language processing (NLP)

that deals with machine reading comprehension.

• Natural language generation (NLG) is a subtopic of natural language processing (NLP) that
is the process of generating text that appears to be written by a human, without the need
for a human to actually write it.

Sources:

• What is natural language processing? Webpage: https://www.ibm.com/topics/natural-

language-processing

• What is machine learning? Webpage: https://www.ibm.com/topics/machine-learning

17
 Natural
language
processing Spam Machine
detection translation
use cases

Chatbots and Social media Text

virtual assistants sentiment analysis summarization

Natural language processing (NLP) is the technology used in many modern applications. The following are a few
examples:

• Spam detection: These technologies use NLP's text classification capabilities to scan emails for language that
often indicates spam or phishing. These indicators can include overuse of financial terms, poor grammar,
threatening language, inappropriate urgency, misspelled company names, and more.

• Machine translation: Useful and effective machine translation involves more than replacing words in one
language with words of another. The process must accurately capture the meaning and tone of the input
language and translate it with the same meaning and desired impact in the output language. IBM Watson
Natural Language Understanding (NLU) and Google Translate are a widely known examples of tools that
provides machine translation.

• Virtual agents and chatbots: Virtual agents, such as Apple's Siri and Amazon's Alexa, use speech recognition to
recognize patterns in voice commands and natural language generation (NLG) to respond with appropriate
actions or comments. Chatbots provide responses to typed text entries (but can also be expanded with voice
interaction, as is the case with IBM Watson Assistant). These technologies can learn to recognize contextual
clues about requests and use them to provide even better responses over time.

• Sentiment analysis: NLP can analyze language used in social media posts, responses, reviews, and more (like
contracts and speeches) to extract attitudes, tones, and emotions in response to products, promotions, and
events–information. Using NLP for sentiment analysis, businesses can uncover hidden data insights from social
media channels to use in product designs, marketing campaigns, and more.

• Text summarization: NLP techniques can digest huge volumes of digital text and create summaries and

18
 synopses for indexes, research databases, or readers who don’t want to read the full text.
Text summarization applications use semantic reasoning and NLG to add useful context
and conclusions to summaries. For example, analyzing a political speech of a company’s
earning reports for key the items discussed, and more.

Notes:

• Natural language processing (NLP) is the technology that gives computers the ability to
understand text and spoken words in much the same way human beings can. NLP
combines computational linguistics — rule-based modeling of human language — with
statistical, machine learning (ML), and deep learning (DL) models. These technologies
enable computers to process human language in the form of text or voice data and to
‘understand’ its full meaning, complete with the speaker or writer’s intent and sentiment.

• Phishing is the fraudulent practice of sending emails purporting to be from reputable

companies to induce individuals to reveal personal information, such as passwords and
credit card numbers.

• Natural language generation (NLG) is a subtopic of natural language processing (NLP) that
is the process of generating text that appears to be written by a human, without the need
for a human to actually write it.

Source:

• What is natural language processing? Webpage: https://www.ibm.com/topics/natural-

language-processing

18
 Conversational
AI

Chatbots and virtual

assistants or agents,
can make it easy for
users to find the
information they need
by responding to their
questions and requests
through text or audio
input without the need
for human intervention,
providing 24-hour
availability to users.

It is expected that the global market for conversational artificial intelligence (AI), including chatbots and intelligent
assistants, will nearly double in the next five years (from 2023 numbers). Conversational AI enables people to
interact with computer applications the way they would with other humans in conversation. Technologies, like
chatbots or virtual agents, use large volumes of data, machine learning (ML), and natural language processing
(NLP) to help imitate human interactions, recognizing speech or text inputs, and translate their meanings across
various languages.

A chatbot is a computer program that simulates and processes human conversation (either written or spoken).
Chatbots, and virtual assistants or agents, can make it easy for users to find the information they need by
responding to their questions and requests through text or audio input without the need for human intervention,
providing 24-hour availability to users. Modern artificial intelligence (AI) chatbots use machine learning (ML), and
natural language processing (NLP) to talk to users as they would talk to a human agent and natural language
understanding (NLU) to discern the user’s need. Deep learning capabilities of AI chatbots are contextually aware
and able to learn as they’re exposed to more and more human language. The longer an AI chatbot has been in
operation, the more detailed and accurate the responses become.

Conversational AI is a cost-effective solution for many business processes and provides quick scalability during
unexpected short-term spikes in demand, such as during a holiday season. The 24-hour availability provides
immediate support that allows users to avoid call center wait times, leading to improvements in the overall user
experience.

The biggest challenge for conversational AI is interpreting the input language (be that language spoken or written).
The human languages’ emotions, tones, and sarcasm make it difficult for this type of technology to interpret
intended user meanings and respond appropriately. User dialects, accents, slang, and more can impact the AI’s
understanding of the raw input, causing problems with processing the input.

19
Notes:

• Machine learning (ML) refers to a broad set of techniques to train a computer to learn from
its inputs, using existing data, and one or more “training” methods, instead of being
explicitly programmed. ML helps a computer to achieve AI.

• Deep learning (DL) is a technique for implementing machine learning (ML) that relies on
deep artificial neural networks to perform complex tasks such as image recognition, object
detection, and natural language processing (NLP). Neural networks are a set of algorithms,
modeled loosely after the neural networks found in the human brain, which are designed to
recognize hidden patterns in data. The “deep” in DL refers to a neural network comprised of
more than three layers (which include the input and the output layers).

• An algorithm is a procedure used for solving a problem or performing a computation.

Algorithms act as an exact list of instructions that conduct a sequence of specified actions
in either hardware- or software-based routines.

• Natural language processing (NLP) is the technology that gives computers the ability to
understand text and spoken words in much the same way human beings can. NLP
combines computational linguistics — rule-based modeling of human language — with
statistical, machine learning (ML), and deep learning (DL) models. These technologies
enable computers to process human language in the form of text or voice data and to
‘understand’ its full meaning, complete with the speaker or writer’s intent and sentiment.

• Natural language understanding (NLU) is a subtopic of natural language processing (NLP)

that deals with machine reading comprehension.

• Neural networks are a set of algorithms, modeled loosely after the neural networks found
in the human brain, that are designed to recognize hidden patterns in data.

Sources:

• What is conversational AI? Webpage: https://www.ibm.com/topics/conversational-ai

• What is natural language processing? Webpage: https://www.ibm.com/topics/natural-

language-processing

• What is a chatbot? Webpage: https://www.ibm.com/topics/chatbots

• What is machine learning? Webpage: https://www.ibm.com/topics/machine-learning

19
 Conversational
AI use cases
Customer support Accessibility HR processes

Healthcare IoT devices Computer software

Conversational artificial intelligence (AI) provides solutions for many business processes and has multiple use
cases across the enterprise. Some examples include the following:

• Online customer support: Online chatbots are replacing human agents along the customer journey. They
answer frequently asked questions (FAQs) around topics, like shipping, provide personalized advice, suggest
products for users, and are changing customer engagements across websites and social media platforms. The
goal here is to create a frictionless experience for clients, but also reduce costs for the business. For example, a
bank teller interaction used to cost (all costs are in USD) a bank about USD 4… today it’s just under a USD 1. An
average Automated Teller Machine (ATM) interaction costs a bank about USD 0.08, but the average interaction
cost for online banking is USD 0.03. This means that every year a bank lagged in technology, it paid
unnecessarily high cost to do business. There are also invisible workflows are previously discussed in this
presentation.

• Accessibility: Companies can become more accessible by reducing entry barriers, particularly for users who
use assistive technologies. Commonly used features of conversational AI for these groups are text-to-speech
dictation and language translation. The auto-generation of closed captioning for millions and millions of videos
on YouTube is a great example of AI in action for accessibility. AI will greatly improve the accessibility space
because AI can help well beyond conversational AI. For example, in the USA, there’s an estimated 3.3
million wheelchair users, with estimated USD 1.825 million of those users over the age of 65 … the number of
senior-aged wheelchair users is increasing every year. Shockingly, 20% of automated wheelchair users
experience an accident ever year (causing damage to themselves and or property) and 11% of them are sent to
a hospital. The wheelchair mobility aid market is a USD 1.14 billion market and computer vision could make
wheelchairs smart and safer.

• Human Resource (HR) processes : The role of human resources (HR) in the modern workplace is more critical

20
 than ever before. As businesses increasingly rely on remote teams and digital tools, HR
professionals must adapt their workflows to manage employees effectively. Luckily, new
technologies like artificial intelligence (AI) and automation are making it easier for HR
professionals to streamline their workflows and make data-driven decisions. Many HR
processes can be optimized by using conversational AI, such as employee training,
onboarding processes, and updating employee information. For example, IBM has fronted
many HR processes with a cognitive assistant (called AskHR) that ‘thinks beyond the bots”.
AskHR can help managers conversationally transfer employees to new jobs or help
employees book time off… fronting the more complex HR Workday system with natural
language interactive workflows. Watson Orchestrate was used by IBM HR to automate
many tasks. Other HR AI use cases include identification of talent for recruitment or
promotion, attrition, customized learning, and more. Learn more at:
https://www.ibm.com/downloads/cas/E1ZJMVAP.

• Healthcare: Conversational AI is making healthcare services more accessible and

affordable for patients, while also improving operational efficiency and administrative
process, such as claim processing. AI can also help with interventions and mental health.
Conversational AI is being used to help treat lonely seniors or determine if an intervention is
required for suicide prevention based on type words. For example, one study found that
usage of the words such as “never”, “everything” , “anymore” , or “always” were 11x more
likely to need an active rescue than a person using the word “suicide”.

• Internet of things (IoT) devices: Most households now have at least one IoT device, from
Amazon Alexa, to smart watches and cell phones. These devices use automated speech
recognition to interact with users. For example, someone can ask Alexa (and IoT device) to
vacuum a room in their house after a spill, and it get the Roomba vacuum (another IoT)
device to start the clean up task; as that person leaves their house, they may ask Alexa to
turn off the light in the house (another IoT device – Apply Home Kit lights) and turn off the
air conditioning (through the IoT Nest device).

• Computer software: Many tasks are simplified by conversational AI, such as search auto
complete on Google, spell check auto-fill, and live caption use in presentations.

Conversational AI helps organizations engage customers, deliver services, and can lower cost
of doing business while increasing profitability.

Sources:

• What is conversational AI? Webpage: https://www.ibm.com/topics/conversational-ai

• Banking costs: Governance in the Digital Age, Canadian Public Policy Forum:
https://ppforum.ca/publications/governance-in-the-digital-age/

20
 Generative AI
applications
Text Code
generation generation

Music Generate Image & video

generation simulated data generation

With the rise of OpenAI’s ChatGPT, generative AI techniques have become popular. This goes far beyond text, and
includes code, simulated data, images, videos, and music. Examples include:

• Text generation: Creative writing can be very fun, yet challenging, especially when someone hit that writer’s
block. Generative AI can generate human sounding text, in massive quantities, and disambiguate (understand
what ‘it’ refers to downstream in the text) that text too. There are loads of uses cases for this. For example,
financial services giant Chase Bank uses artificial intelligence (AI) to write marketing copy after a pilot program
saw click-through rates on AI-written ads jump by as much as 450%. In the case of one of their promotions, AI
generated “LIMITED TIME OFFER: We’ll reward you with $5 Cash Back when you go paperless” and that got
nearly 2x the unique clicks that the human copywriters’ message “Go paperless and earn $5 Cash Back”. But
the bank was quick to note that that eliminating humans from the marketing copywriting process is not the
goal, rather to get copy that resonates. Edits and reviews have always been integral to the process… and
humans are in the loop. It’s not humankind versus machine… it’s human-kind aided by machine.

• Code generation: Generates code based on user requests. Ideal for boilerplate and simple scenarios, allowing
the user to focus on the more difficult work. IBM Watson Code Assistant (previously known as project
“Wisdom”) has learned the language of Ansible Playbooks and can generate code based on a text prompt; for
example, “Write me a Ansible playbook to deploy Node.js stack on my developer servers.”

• Music generation: Write new melodies and songs without any music skills. Simulate what new songs could
sound like with different instruments and tones. For example, a piano cover refers to a musical piece in which
all musical elements of an existing song are recreated or arranged in the form of a piano performance. Piano
covers of pop music are one of the widely enjoyed forms of music or background ambient noise (like hold
music). People also use piano covers for music education purpose and general entertainment; in fact, piano
cover creators have millions of subscribers on media such as YouTube. For humans to create piano covers, it is
necessary to recognize all musical elements such as melodies, chords, or moods from the original audio and

21
 reinterpret them into musically appropriate piano performances. Therefore, making a piano
cover is not an easy task even for humans as it is a creative task and requires musical
knowledge. The Pop2Piano deep learning (DL) model is an AI that has shown excellent
performance in modeling high-dimensional music audio data. Pop2Piano is a transformer-
based piano cover generation that generates a piano performance (MIDI) from the
waveform of pop music. Its training data set was built on on 5989 piano covers from 21
arrangers and corresponding pop songs on YouTube for a total for 307 hours of music.

• Simulated (synthetic) data generation: When larger sample sizes of data are needed,
foundation models can generate data that is like the existing data sets. Specifically,
synthetic data is information that's been generated on a computer to augment or replace
real data to improve AI models, protect sensitive data, and mitigate bias. Synthetic data can
be used for hackathons, product demos, and internal prototyping to replicate a set of data
with the right statistical attributes. For example, banks and financial services institutions
use synthetic data by setting up multiagent simulations to explore market behaviors (such
as pension investments and loans), to make better lending decisions or to combat financial
fraud. Retailers use synthetic data for autonomous check-out systems, cashier-less stores,
or analysis of customer demographics. Ironically, data generated by AI and can help AI!
Synthetic data can be used to increase the accuracy of machine learning (ML) models. Real-
world data is happenstance and does not contain all permutations of conditions or events
possible in the real world. Synthetic data can counter this by generating data at the edges,
or for conditions not yet seen. The breadth of synthetic data’s applicability will make it a
critical accelerator for AI. Synthetic data makes AI possible where lack of data makes AI
unusable due to bias or inability to recognize rare or unprecedented scenarios.

• Image and video generation: Enable people without graphic design or artistic skills to
create high quality custom images and reduce dependance on stock imagery. This is seen
by generative networks such as Stable Diffusion, DALLE-2, Midjourney, NightCafe, and
more. Simple prompts like “a polar bear playing the base guitar on an iceberg floating by a
tropical island in an oil panting setting that looks like the styling of VanGogh” will produce a
new unique image along these instructions. But AI can help in other areas beyond the
creative that are more applicable to everyday life. For example, AI can make realistic edits
to existing images and add and remove elements while taking shadows, reflections, and
textures into account. Photoshop, Topaz Gigapixel AI, Pixelmator (and other image tools)
embed this kind of AI today by filling in pixels for removed objects (like someone’s hand in
an almost perfect shot) or upscaling the quality of an image. Outpainting is another
example of AI working on images; in this use case images are expanded beyond what’s in
the original canvas, creating expansive new compositions. For example, extended a beach
shore with ‘more of the same’ to add other items to a broadened picture.

Notes:

• Generative AI refers to a set of AI algorithms that can generate new outputs — such as
text, images, code, or audio — based on the training data, unlike traditional AI systems that
are designed to recognize patterns and make predictions. Sometimes the AI that powers
these solutions are referred to as decoders.

21
• Deep learning (DL) is a technique for implementing machine learning (ML) that relies on
deep artificial neural networks to perform complex tasks such as image recognition, object
detection, and natural language processing (NLP). Neural networks are a set of algorithms,
modeled loosely after the neural networks found in the human brain, which are designed to
recognize hidden patterns in data. The “deep” in DL refers to a neural network comprised of
more than three layers (which include the input and the output layers).

• Machine learning (ML) refers to a broad set of techniques to train a computer to learn from
its inputs, using existing data, and one or more “training” methods, instead of being
explicitly programmed. ML helps a computer to achieve AI.

• Neural networks are a set of algorithms, modeled loosely after the neural networks found
in the human brain, that are designed to recognize hidden patterns in data.

• Natural language processing (NLP) is the technology that gives computers the ability to
understand text and spoken words in much the same way human beings can. NLP
combines computational linguistics — rule-based modeling of human language — with
statistical, machine learning (ML), and deep learning (DL) models. These technologies
enable computers to process human language in the form of text or voice data and to
‘understand’ its full meaning, complete with the speaker or writer’s intent and sentiment.

• An algorithm is a procedure used for solving a problem or performing a computation.

Algorithms act as an exact list of instructions that conduct a sequence of specified actions
in either hardware- or software-based routines.

• Foundation models are typically built using a specific kind of neural network architecture,
called a transformer, which is designed to generate sequences of related data elements (for
example, like a sentence).

• Bias is a type of data error that can occur if an artificial intelligence (AI) models’ output is
skewed by the model’s training data. For example, a model may associate specific traits or
professions with a certain race or gender, leading to inaccurate predictions and offensive
responses.

• A transformer model is a neural network architecture useful for understanding language,

which does not have to understand words one at a time but can look at an entire sentence
at once. A technique called self-attention allows the model to focus on the particular words
that are important in understanding the meaning of the sentence. Transformers single-
handedly spawned the world of foundation models and its birthplace research paper
(Attention is all You) spawned the AI inflection point that is happening today. AI
transformer models can be Encoder models (like BERT), Decoder models (like GPT), or
Encoder and Decoder models (like T5). Learn more about Transformers at:
https://www.youtube.com/watch?v=ZXiruGOCn9s.

Sources:

21
• What is artificial intelligence? Webpage: https://www.ibm.com/topics/artificial-intelligence

• Wall Street Journal article, JP Morgan Chase Taps AI to Make Marketing Messages More
Powerful: https://www.wsj.com/articles/jpmorgan-chase-taps-ai-to-make-marketing-
messages-more-powerful-11564482606

21
 Glossary of
• Algorithm
AI terms •
•
Artificial general intelligence (AGI)
Artificial intelligence (AI)
• Artificial narrow intelligence (ANI)
• Artificial super intelligence (ASI)
• Bias
• Convolutional neural network (CNN)
• Deep learning (DL)
• Foundation model
• Generative AI
• Hallucination
• Large language model (LLM)
• Machine learning (ML)
• Natural language generation (NLG)
• Natural language processing (NLP)
• Natural language understanding (NLU)
• Neural networks
• Parameters
• Transformer model

This glossary of terms related to artificial intelligence (AI) is provided as a summary to use for reference.

• An algorithm is a procedure used for solving a problem or performing a computation. Algorithms act as an
exact list of instructions that conduct a sequence of specified actions in either hardware- or software-based
routines.

• Artificial general intelligence (AGI), general AI, is a theoretical form of AI where a machine would have an
intelligence equaled to humans; it would have a self-aware consciousness that can solve problems, learn, and
plan for the future. This is also known as strong AI.

• Artificial intelligence (AI) is the field focused on building smart machines that can perform tasks that normally
require human intelligence.

• Artificial narrow intelligence (ANI), narrow AI, also known as weak AI, is AI trained and focused to perform
specific tasks. This AI drives most of the AI that surrounds us today, such as Apple's Siri, Amazon's Alexa, IBM
Watson, and autonomous vehicles.

• Artificial super intelligence (ASI), super AI, also known as superintelligence, would surpass the intelligence
and ability of the human brain. While strong AI is still entirely theoretical with no practical examples in use
today, that doesn’t mean AI researchers aren't also exploring its development.

• Bias is a type of data error that can occur if an artificial intelligence (AI) models’ output is skewed by the
model’s training data. For example, a model may associate specific traits or professions with a certain race or
gender, leading to inaccurate predictions and offensive responses.

22
• Convolutional neural network (CNN) is a network architecture for deep learning
specifically used for image recognition and tasks that involve the processing of pixel data.
CNNs are particularly useful for finding patterns in images to recognize objects, classes,
and categories.

• Deep learning (DL) is a technique for implementing machine learning (ML) that relies on
deep artificial neural networks to perform complex tasks such as image recognition, object
detection, and natural language processing (NLP). Neural networks are a set of algorithms,
modeled loosely after the neural networks found in the human brain, which are designed to
recognize hidden patterns in data. The “deep” in DL refers to a neural network comprised of
more than three layers (which include the input and the output layers).

• Foundation models are typically built using a specific kind of neural network architecture,
called a transformer, which is designed to generate sequences of related data elements (for
example, like a sentence).

• Generative AI refers to a set of AI algorithms that can generate new outputs — such as
text, images, code, or audio — based on the training data, unlike traditional AI systems that
are designed to recognize patterns and make predictions. Sometimes the AI that powers
these solutions are referred to as decoders.

• Hallucination is a well-known phenomenon in large language models (LLMs) in which the

system provides an answer that is factually incorrect, irrelevant, or nonsensical because of
limitations in it’s training data and architecture; more concerning is the hallucinated answer
sounds plausible.

• A large language model (LLM) is a type of machine learning model that has been trained
on large quantities of unlabeled text using self-supervised learning and can perform a
variety of natural language processing (NLP) tasks (even when that language is a
programming language). Output may range from books, articles, social media posts, online
conversations, and even code. The architecture of an LLM consists of layers of neural
networks that learn to generate language in a way that is similar to how humans use
language.

• Machine learning (ML) refers to a broad set of techniques to train a computer to learn from
its inputs, using existing data, and one or more “training” methods, instead of being
explicitly programmed. ML helps a computer to achieve AI.

• Natural language generation (NLG) is a subtopic of natural language processing (NLP) that
is the process of generating text that appears to be written by a human, without the need
for a human to actually write it.

• Natural language processing (NLP) is the technology that gives computers the ability to
understand text and spoken words in much the same way human beings can. NLP
combines computational linguistics — rule-based modeling of human language — with
statistical, machine learning (ML), and deep learning (DL) models. These technologies

22
 enable computers to process human language in the form of text or voice data and to
‘understand’ its full meaning, complete with the speaker or writer’s intent and sentiment.

• Natural language understanding (NLU) is a subtopic of natural language processing (NLP)

that deals with machine reading comprehension.

• Neural networks are a set of algorithms, modeled loosely after the neural networks found
in the human brain, that are designed to recognize hidden patterns in data.

• Parameters are numerical values that define a large language model’s (LLM) structure and
behavior, like clues that help it guess what words come next. Modern systems like ChatGPT
are thought to have hundreds of billions of parameters.

• A transformer model is a neural network architecture useful for understanding language,

which does not have to understand words one at a time but can look at an entire sentence
at once for context and disambiguation. A technique called self-attention allows these
models to focus on the particular words that are important in understanding the meaning of
the sentence. Transformers single-handedly spawned the world of foundation models and
its birthplace research paper (Attention is all You Need) spawned the AI inflection point
that is happening today. AI transformer models can be Encoder models (like BERT),
Decoder models (like GPT), or Encoder and Decoder models (like T5). Learn more about
Transformers at: https://www.youtube.com/watch?v=ZXiruGOCn9s.

22
These videos serve as a complement to the learning in in this presentation

Resources

Review these videos to • AI vs Machine Learning

continue learning about • Natural Language Processing vs Natural
the concepts introduced
in this presentation. Language Understanding and Generation
• What are Generative AI models?
• What are Transformers (Machine Learning
Model)?
• Risks of Large Language Models (LLM)
• Why Large Language Models Hallucinate
• Artificial Intelligence – Are We There Yet?

This self-explanatory slide lists additional resources to review.

There are high-quality informative videos on many of the AI technologies being applied to IBM offerings. These
videos serve as a complement to the learning in in this presentation and can be great assets to share with clients
who may be equally curious about these foundational topics.

Resources:

• AI vs Machine Learning:
https://www.youtube.com/watch?v=4RixMPF4xis&list=PLOspHqNVtKADfxkuDuHduUkDExBpEt3DF

• Natural Language Processing vs Natural Language Understanding and Natural Language Generation:
https://www.youtube.com/watch?v=1I6bQ12VxV0&list=PLOspHqNVtKADfxkuDuHduUkDExBpEt3DF

• What are Generative AI models?:

https://www.youtube.com/watch?v=hfIUstzHs9A&list=PLOspHqNVtKADfxkuDuHduUkDExBpEt3DF

• What are Transformers (Machine Learning Model)?:

https://www.youtube.com/watch?v=ZXiruGOCn9s&list=PLOspHqNVtKADfxkuDuHduUkDExBpEt3DF

• Risks of Large Language Models (LLM):

https://www.youtube.com/watch?v=r4kButlDLUc&list=PLOspHqNVtKADfxkuDuHduUkDExBpEt3DF

• Why Large Language Models Hallucinate: https://www.youtube.com/watch?v=cfqtFvWOfg0

• Artificial Intelligence – Are We There Yet?: https://www.youtube.com/watch?v=hdKYPxzeerU

23
Notes:

• Machine learning (ML) refers to a broad set of techniques to train a computer to learn from
its inputs, using existing data, and one or more “training” methods, instead of being
explicitly programmed. ML helps a computer to achieve AI.

• Natural language generation (NLG) is a subtopic of natural language processing (NLP) that
is the process of generating text that appears to be written by a human, without the need
for a human to actually write it.

• Natural language processing (NLP) is the technology that gives computers the ability to
understand text and spoken words in much the same way human beings can. NLP
combines computational linguistics — rule-based modeling of human language — with
statistical, machine learning (ML), and deep learning (DL) models. These technologies
enable computers to process human language in the form of text or voice data and to
‘understand’ its full meaning, complete with the speaker or writer’s intent and sentiment.

• Natural language understanding (NLU) is a subtopic of natural language processing (NLP)

that deals with machine reading comprehension.

• Generative AI refers to a set of AI algorithms that can generate new outputs — such as
text, images, code, or audio — based on the training data, unlike traditional AI systems that
are designed to recognize patterns and make predictions. Sometimes the AI that powers
these solutions are referred to as decoders.

• An algorithm is a procedure used for solving a problem or performing a computation.

Algorithms act as an exact list of instructions that conduct a sequence of specified actions
in either hardware- or software-based routines.

• A transformer model is a neural network architecture useful for understanding language,

which does not have to understand words one at a time but can look at an entire sentence
at once. A technique called self-attention allows the model to focus on the particular words
that are important in understanding the meaning of the sentence. Transformers single-
handedly spawned the world of foundation models and its birthplace research paper
(Attention is all You) spawned the AI inflection point that is happening today. AI
transformer models can be Encoder models (like BERT), Decoder models (like GPT), or
Encoder and Decoder models (like T5). Learn more about Transformers at:
https://www.youtube.com/watch?v=ZXiruGOCn9s.

• Neural networks are a set of algorithms, modeled loosely after the neural networks found
in the human brain, that are designed to recognize hidden patterns in data.

• Foundation models are typically built using a specific kind of neural network architecture,
called a transformer, which is designed to generate sequences of related data elements (for
example, like a sentence).

23
• A large language model (LLM) is a type of machine learning model that has been trained
on large quantities of unlabeled text using self-supervised learning and can perform a
variety of natural language processing (NLP) tasks (even when that language is a
programming language). Output may range from books, articles, social media posts, online
conversations, and even code. The architecture of an LLM consists of layers of neural
networks that learn to generate language in a way that is similar to how humans use
language.

23
 Let’s get
started!

In summary, this presentation was designed to introduce the basic concepts of artificial intelligence (AI), it’s use
cases, and some of the terminology that is frequently used in discussions about AI. This basic understanding of AI
and associated concepts, will prepare you for more fluent conversation with clients about AI.

24
This self-explanatory slide does not require speaker notes.

25