Quick Reference

Guide for
Understanding AI
This Quick Reference Guide for AI is designed to be your
go-to resource for understanding key principles, best
practices, and technologies.
By: Ian Horowitz

1
Quick Reference Guide for Understanding AI

Introdution
Artificial intelligence (AI) is the superstar of the tech world today, with virtually
every industry considering how AI can enhance products, services, and
competitive strengths. At the same time, discussions about the ethical use of AI
have led to persistent concerns and the creation of new policies and laws.
While initial impressions of AI may evoke a mix of excitement, skepticism, worry,
and countless questions, it’s important to recognize why AI matters. AI aims to
replicate human-like cognitive abilities through computational models and
underlying algorithms. By providing machines with the power to learn, adapt, and
perform tasks that usually require human intellect, AI represents a transformative
era in computing and IT.

This Quick Reference Guide is designed to assist you in exploring fundamental AI

principles, best practices, and cutting-edge technological developments. Inside,
you’ll find essential AI terminology and clear explainers to help you start your AI
journey.

2
Quick Reference Guide for Understanding AI

Table of Contents
1. Types of AI
Narrow AI vs. General AI / Machine Learning (ML) / Deep Learning (DL)
Natural Language Processing (NLP) / Large Language Model (LLM)

2. How Is AI Used Today?

Financial Services / Healthcare / Retail / Manufacturing and Industrial

3. The Field of AI Ethics

Responsible AI / Fairness and Bias / Privacy

4. AI Frameworks
TensorFlow / PyTorch / scikit-learn

5. AI Security

6. Regulations and Compliance

3
Quick Reference Guide for Understanding AI

1. Types of AI
Narrow vs. General AI

When we discuss AI, we refer to two overarching types of systems: Narrow AI

and General AI.

Narrow AI (also known as "Weak AI") is designed and trained for specific tasks. It
operates within a limited context and awareness, which is typical of most AI
systems today. Narrow AI excels in executing singular tasks such as facial
recognition or language translation but can’t extend its intelligence beyond its
programmed scope. Examples include AI-powered search engines,
recommendation algorithms, chatbots, and voice assistants like Apple’s Siri and
Amazon’s Alexa.

In contrast, General AI (also referred to as “Strong AI”) is a theoretical concept of

intelligent systems that could potentially understand, learn, and apply machine
intelligence across a broad spectrum of tasks, displaying human-like or even
transhuman cognitive abilities. In theory, General AI could possess or
demonstrate some level of sentience, self-awareness, emotion, and an
understanding of abstract concepts.

Machine Learning (ML)

Machine Learning (ML) is a subset of AI based on enabling machines to do two

things:

1. Learn from inputs.

2. Use pattern recognition to make complex inferences and decisions.

4
Quick Reference Guide for Understanding AI

ML operates on underlying algorithms that get better over time through learned
experiences and training data.

ML is broadly classified into three main types:

1. Supervised Learning: The algorithm is shown a lot of examples (i.e., labeled

datasets) of how it should “think” until it can make accurate predictions. As
the algorithm recognizes patterns and relationships in the data it’s shown, it
adjusts its internal parameters to make its predictions increasingly accurate.
When the algorithm makes mistakes, it can be given feedback to improve its
generalizations.
2. Unsupervised Learning: The algorithm is given a lot of unlabeled data and
made to identify patterns, connections, or structures within the data without
any clear external guidance. Unsupervised learning is commonly used for vast
datasets because it can discover relationships that would otherwise be
overlooked through manual assessments.
3. Reinforcement Learning: An algorithm or entity (called an “agent”) learns by
receiving rewards or penalties as it explores an environment. The agent uses
trial and error and is given a reward signal or penalty signal based on its
actions. The agent then adjusts its strategy based on this feedback.

ML has a variety of applications.

Examples:
Recommendation engines used by companies including Amazon, Meta, and
Netflix.
Records processing and diagnostic assistance in AI medical tools, such as
those spun off by IBM’s failed Watson health initiative.

Deep Learning

Deep learning (DL) is a subfield of machine learning. Deep learning relies on

multi-layer neural networks, including Artificial Neural Networks (ANNs),

5
Quick Reference Guide for Understanding AI

Convolutional Neural Networks (CNNs), and Recurrent Neural Networks (RNNs).

Neural networks enable machines to process data in nuanced, almost human-like
ways.

DL models excel at identifying and interpreting elements from basic, unprocessed

data. This makes them adept at performing complex tasks like image and speech
recognition.

A quintessential application of deep learning today is in the field of computer

vision (CV), where it enables machines to interpret and understand the content of
images and videos. This technology plays a key role in the ability of autonomous
vehicles to "see" and navigate safely. Current production models of Tesla
vehicles use neural networks, which are largely reliant on GPUs from AI hardware
and software vendor Nvidia.

Natural Language Processing (NLP)

Natural language processing (NLP) is a part of machine learning that sometimes

incorporates deep learning models. NLP focuses on the interaction between
computers and human language, teaching machines to understand, interpret, and
respond to language.

A primary application of NLP is language translation services, such as those seen

in Google Cloud’s translation tools. In addition, NLP, along with deep Learning,
has been important in the development of sophisticated Large Language Models
(see the next section). These models are behind popular chat-based tools such
as OpenAI's ChatGPT and Google's Gemini.

Large Language Model (LLM)

Large Language Models (LLMs) are AI systems designed to understand,

generate, and interact using human language. These models are trained on vast
datasets of text, enabling them to grasp language patterns, nuances, and
contexts. Text generation models, like those developed by OpenAI, generate text
based on the input they receive.
6
Quick Reference Guide for Understanding AI

The applications of LLMs range from automating email responses and

organizational office tasks to more complex tasks such as writing code.

Generative AI

Generative AI, or GenAI, is a type of AI system that uses algorithms to generate

content. GenAI systems today can produce content in a variety of mediums,
including the following:

Text – prose or verse in basically any format and genre, including nonfiction
and fiction.
Pictures – photorealistic and creative imagery in imitations of artistic styles.
Video – videos and animations.
Audio – imitations of human voices as well as music composition.

Microsoft’s Copilot is an example of a system that uses GenAI and LLMs to

generate personalized assistance to users. For example, Copilot is integrated
with various Microsoft 365 apps, such as Word and Excel, to help with everyday
work tasks.

GenAI can also be used to generate data. For example, GenAI can create
synthetic data to help fill out datasets for use in machine learning models.

Many fields are currently using GenAI, such as pharmaceuticals, where it can help
create molecular structures for new drugs.

You will learn much more about GenAI throughout this guide.

7
Quick Reference Guide for Understanding AI

2. How Is AI Used Today?

Let’s look at how several major industries are applying and testing AI
technologies.

Financial Services

In the banking and financial services sector, AI has become instrumental in

detecting potential fraud and suspicious activity. Machine learning algorithms, in
particular, can be useful in fraud prevention. Companies like PinDrop use machine
learning for user verification and fraud prevention in the banking sector.

Healthcare

Despite IBM’s AI incursions into the healthcare industry not paying off, AI is
making a significant impact in drug discovery and diagnostics. AI can expedite
the drug development process by analyzing large datasets, predicting molecular
interactions, and simulating drug responses. This allows for faster identification
of potential drug candidates and the development of new therapies.

From big tech companies to smaller startups, AI-driven drug development is

widely seen as the catalyst for a coming revolution in healthcare. Interest is
reflected in substantial venture capital funding, reaching $2.1 billion in the first
half of 2021 alone. Google's DeepMind subsidiary has also been reported to have
made large investments in this type of technology.

Retail

AI has a growing role in delivering personalized shopping recommendations.

Various retailers and online marketplaces, most notably Amazon, rely on machine
learning algorithms to push recommendations to their customers.

8
Quick Reference Guide for Understanding AI

Recommendations powered by AWS' Amazon Personalize service are tailored to

users' preferences based on their interactions, purchase history, and browsing
behavior. Amazon Personalize gives retailers and service providers access to the
same ML algorithm used by Amazon.com’s recommendation system.

Although not "retail" businesses in a traditional sense, subscription-based media

and streaming services, including Spotify, rely heavily on machine learning. They
deploy sophisticated recommendation algorithms to improve subscriber
engagement and retention by anticipating users' interests.

Manufacturing and Industrial

In manufacturing, AI has proven vital in improving efficiency and reducing

downtime. In addition to automation, a common AI application is predictive
equipment maintenance. Machine learning algorithms enable factories to predict
when equipment might fail or need maintenance.

Major industry players, such as General Electric, Honeywell, Siemens, and IBM,
offer proprietary versions of AI-powered technology for predictive maintenance.
The space is also filled with smaller competitors offering custom products,
ranging from IoT monitoring and service log maintenance to AI-based warning
systems.

9
Quick Reference Guide for Understanding AI

3. The Field of AI Ethics

Responsible AI

Incorporating ethical principles into the design and development of AI systems is

necessary to mitigate biased decisions. In discussions about Responsible AI, an
AI framework that focuses on ethical and legal interests, you will come across
concepts and principles such as the ones we define here (in very simple terms):

Data bias – The datasets used to train AI models are biased and create an
incomplete picture because they lack diverse and representative information.
Inherent data bias – Bias is ingrained into the datasets used to train AI
models due to the nature of the dataset itself.
Fairness – In an AI context, fairness is the assurance that AI systems don’t
favor one group over another and are free from biases in their decision-
making processes.
Transparency – Transparent AI systems means that they are open to scrutiny
and capable of being explained by AI developers.
Explainable AI (XAI) – Explainable AI describes AI systems that can
demonstrate how they arrive at a decision. In contrast, “black box” AI systems
can’t show how they make decisions.
Accountability – In an AI context, accountability means AI developers and
users take responsibility for the impact of an AI system and its decisions.

Regulatory frameworks and guidelines for AI deployment are currently in

development. The European Commission has proposed a legal framework for AI,
which includes clear requirements for high-risk AI systems, suggested market
regulations, and stringent enforcement measures.

In the U.S., although there is no comprehensive federal legislation governing the

sale, marketing, or use of AI, various initiatives and guidelines encourage the
adoption of explainable AI over black-box AI models.

10
Quick Reference Guide for Understanding AI

Deep learning is frequently described as a black-box AI model because it tends

to lack interpretability and explainability. In fact, complex deep learning and
neural network architectures can be so opaque that computer scientists and
researchers may struggle to understand how their systems work and make
decisions.

The National Institute of Standards and Technology (NIST) has developed a set
of principles for explainable AI, aiming to address concerns related to
interpretability and explainability deficits seen in some AI systems.

Fairness and Bias

AI is increasingly handling complex decision-making tasks. As such, bias must be

minimized to ensure fairness and equity. However, weeding out bias in AI
represents a huge challenge. As AI systems grow more intricate, interpretability
becomes more difficult. This highlights the need for maintaining accountability
through the responsible design and deployment of AI systems.

Privacy

Many people inside and outside of the tech industry agree that it’s incredibly
important for AI systems to respect the fundamental rights, privacy, and
autonomy of users. To ensure this, the development and use of AI should be
guided by stringent safety measures, incorporating human-based monitoring and
auditing processes to reduce potential risks.

To monitor AI technologies for privacy breaches, organizations must implement

testing procedures, external validation, and ongoing oversight. This becomes
critical when operating in markets outside the U.S., where data protections are
stronger – e.g., in the European Union, which has the General Data Protect
Regulation (GDPR) laws.

GDPR governs data use and offers individuals significant control over personal
data. These protections affect how AI systems can be designed and used in the
EU. Companies deploying AI systems internationally must also be aware that
11
Quick Reference Guide for Understanding AI

individual countries have their own data privacy policies. For example, Italy
temporarily blocked access to ChatGPT in 2023 due to its data privacy policies

4. AI Frameworks
TensorFlow

Developed by Google, TensorFlow is an open-source framework widely used by

AI developers for its flexibility and comprehensive toolkit.

TensorFlow is well-suited for large-scale, complex computations and offers

extensive support for neural networks. Many deep learning applications are built
with the TensorFlow framework.

TensorFlow is popular in both research and production settings, excelling in

areas such as speech and image recognition and predictive analytics.

PyTorch

PyTorch, created by Facebook's AI Research lab, has gained popularity for its
user-friendly interface and dynamic computational graph. It's particularly favored
in the academic and research community due to its straightforward way of
creating and experimenting with neural networks. PyTorch’s flexibility allows
developers to make on-the-fly changes to complex network architectures.

PyTorch is part of the Linux Foundation’s collection of open-source software for

developers.

scikit-learn

scikit-learn is mainly used for traditional machine learning algorithms. Built on top
of SciPy (Scientific Python), its data mining and analysis tools make it a popular
choice for ML-dependent applications. Applications include processing large
datasets, automating industrial processes, and other routine activities where a
high level of explainability is desired.

12
Quick Reference Guide for Understanding AI

5. AI Security
AI needs data for training, validation, and operation. As such, data protection a
top priority.

Similar to any server or cloud-based technology, AI systems can have

vulnerabilities that leave it exposed to cyberattacks and data breaches. In
addition to safeguarding user data, IT professionals must vigilantly protect AI
systems from malicious manipulation, which could lead to harmful outcomes. For
example, subtle changes in input data by internal or external actors can lead an
AI system to produce erroneous or biased outputs.

The security of AI systems requires a multi-layered approach, integrating

advanced encryption and secure data storage technologies. It also requires
stringent multifactor authorizations and access protocols.

6. Regulations and Compliance

While we’ve already discussed the GDPR and the EU's proposed legislation on AI,
for IT professionals operating in the U.S. market, it's a good idea to familiarize
yourself with the California Consumer Privacy Act (CCPA).

The CCPA deals primarily with consumer data rights. It ensures certain
protections for how data can be handled and sold, also specifying that
companies must disclose how they use data. For any AI system that processes
large amounts of U.S. consumer data, it’s important to fully comply with with the
CCPA. Additional U.S. regulations are likely ahead.

13
Quick Reference Guide for Understanding AI

About ITPro Today

IProToday.com is a leading online source of daily news, analysis, opinions and
how-to’s about the information technology industry. Along with offering practical
IT operations and career insights, we help IT professionals and technology
stakeholders learn about, assess and manage the acquisition of next-gen
technology that drives business innovation, including — but not limited to —
analytics, artificial intelligence/machine learning, cloud computing, low-code/no-
code, DevOps, NoOps, DataOps, compute engines, containers, edge computing,
hyper-converged infrastructure, security, software development and storage.

14