A Beginner’s Guide to

Machine Learning
Fundamentals

udacity.com
A B e g i n n e r ’ s G u i d e to M ac h i n e L e a r n i n g F u n da m e n ta l s

It was only a couple of decades back that, to many of us, the idea of programming machines to execute
complex, human-level tasks seemed as far away as the science fiction galaxies these technologies could have
emerged from. Fast-forward to today, and the field of machine learning reigns supreme as one of the most
fascinating industries one can get involved in. From its breakneck pace of innovation to its real-time cultural
impact, machine learning is a line of work that isn’t for the faint of heart. It’s one that rewards the curious,
favors the bold, and will go only as far as the imaginations of the professionals who run it. And chances are, if
you clicked on this article, those are the exact things that light you up about the industry.

Here at Udacity, machine learning is in our DNA – it’s a field we’re proud to have been on the ground floor of. In
this guide, we’ll cover fundamental concepts and definitions of the technology, how it works, frequently asked
questions about the space, and how you can get involved in shaping it. Let’s get started!

Table of Contents
The Evolution of Machine Learning

Types of Machine Learning

Machine Learning vs. Traditional Programming

How Machine Learning Works

Machine Learning FAQs

Machine Learning, Defined

Machine Learning is…
Machine learning (ML) is a subfield of artificial intelligence that empowers computers to learn and make predictions
or decisions without being explicitly programmed. In simpler terms, it’s a set of techniques that allows computers to
analyze data, recognize patterns, and continuously improve their performance. This enables these machines to
tackle complex tasks that were once reserved for human intelligence only, like image recognition, language
translation, and even helping cars drive autonomously.

udacity.com
A B e g i n n e r ’ s G u i d e to M ac h i n e L e a r n i n g F u n da m e n ta l s

The Importance of Machine Learning

The importance of machine learning in today’s technology landscape can’t be overstated. It has quickly become a
part of our daily lives, impacting a wide variety of industries – from healthcare and finance to entertainment and
transportation. Here are just a few key reasons why, in the context of the products we enjoy and the technology
that powers them, ML matters:

Automation: ML automates repetitive and time-consuming tasks, increasing efficiency and reducing human error.

Personalization: ML algorithms power recommendation systems on platforms like Netflix and Amazon, tailoring content and
products to individual preferences.

Data-driven Insights: ML extracts valuable insights from massive datasets, aiding decision-making and strategy formulation.

Innovation: ML is driving innovation in areas like autonomous vehicles, healthcare diagnostics, and natural language processing.

The Evolution of Machine Learning

19 5 0s-19 6 0s: The Birth of ML
The roots of ML can be traced back to the 1950s and 1960s when pioneers like Alan Turing and Arthur Samuel laid the
groundwork for the technology. Turing introduced the concept of a “learning machine,” while Samuel developed the first
self-learning program to play checkers, a groundbreaking moment in ML history.

1970s-1980s: Rule-Based Systems

During this period, ML largely relied on rule-based systems. Expert systems, which codified human knowledge into rules,
were prevalent. Though effective in some instances, they struggled with complex, real-world problems.

1990s: Emergence of Neural Networks

The 1990s witnessed a resurgence of interest in “neural networks” (a machine learning model inspired by the human
brain’s structure, used to solve complex tasks like natural language processing). Researchers developed backpropagation
algorithms, which allowed neural networks to learn and generalize from data. However, computational limitations
hampered their progress.

2000s: Big Data and Improved Algorithms

The 2000s witnessed a number of milestones that propelled ML forward. The introduction of “big data” provided ample
data for machines to be trained on – while improved algorithms and computing power made deep learning and complex
models practical. This era marked the beginning of ML’s modern renaissance.

2010s: Deep Learning Revolution

The 2010s were dominated by the deep learning revolution. “Deep neural networks” achieved remarkable success in
image and speech recognition. Defined, this is a type of artificial neural network with multiple hidden layers between the
input and output, enabling it to model complex relationships and extract hierarchical features from data. Key
breakthroughs like the AlexNet and AlphaGo demonstrated the power of deep learning.

2020s and Beyond: Widening Applications

As we enter the 2020s, ML is becoming increasingly integrated into everyday life. It plays a crucial role in finance,
healthcare, autonomous vehicles, and many other fields. With the development of explainable AI and ethical
considerations, ML continues to evolve, addressing challenges related to transparency around how a model and
algorithm makes decisions, and fairness to ensure models don’t discriminate against individuals based on underlying
characteristics like race, gender, age, religion, and socioeconomic status.

udacity.com
A B e g i n n e r ’ s G u i d e to M ac h i n e L e a r n i n g F u n da m e n ta l s

Types of Machine Learning

To better understand the technology at a high level, we’re going to dive into the three main types of machine
learning – along with their real-world applications, advantages, and disadvantages.

Supervised Learning
U nderstanding Su p erv ised Learning

Supervised learning is one of the foundational paradigms in machine learning. In this approach, the algorithm
learns from a labeled dataset, which means the input data is paired with the correct output or target. The idea is for
the algorithm to map input data to the correct output based on the patterns it learns during training.

Some real-world examples of supervised learning include:

Image Classification: Identifying objects in images (“cat” vs. “dog”).

NaturalLanguage Processing (NLP): Language translation,

sentiment analysis, and virtual assistants.

Medical Diagnosis: Detecting diseases from medical images or

patient data.

Email Filtering: Classifying emails as spam or not.

Recommendation Systems: Suggesting products or content based

on user behavior.

Autonomous Vehicles: Recognizing road signs and pedestrians.

Advantages and Limitations o f Su p erv ised Learning

Advantages:
Accurate Predictions: Supervised models are capable of making highly accurate predictions due to their labeled training data.

Interpretability: Models can provide insights into why a certain prediction was made.

Limitations:
Labeled Data Requirement: It depends on labeled data for training, which can be expensive and time-consuming to obtain.

Limited Generalization: Models might struggle with data outside their training scope.

Bias: If the training data is biased, the model can inherit those biases.

Unsupervised Learning
W hat is U nsu p er v ised Learning ?

Unsupervised learning is a branch of machine learning where the algorithm works with unlabeled data. Unlike
supervised learning, this type doesn’t have specific target outputs. Instead, it seeks to discover hidden patterns or
structures within the data.

Some real-world examples of supervised learning include:

Clustering: Grouping similar data points together.

Dimensionality Reduction: Reducing the complexity of data while preserving important information.

Customer Segmentation: Identifying groups of customers with similar buying behavior.

Anomaly Detection: Detecting fraudulent transactions in financial data.

Topic Modeling: Extracting themes from a collection of documents.

udacity.com
A B e g i n n e r ’ s G u i d e to M ac h i n e L e a r n i n g F u n da m e n ta l s

Advantages and Limitations of Unsu p erv ised Learning

Advantages:
Discovering Hidden Patterns: Unsupervised learning is excellent at identifying hidden structures within data that might not be
apparent through manual inspection – which is valuable for data exploration and gaining insights.

Data Dimensionality Reduction: More advanced techniques like Principal Component Analysis (PCA) and t-SNE (t-Distributed
Stochastic Neighbor Embedding) can reduce the dimensionality of high-dimensional data, making it more manageable for analysis
and visualization.

Limitations:
Lack of Clear Objectives: Unsupervised learning often lacks clear objectives or specific goals. It can be challenging to evaluate the
success of an unsupervised learning model because there may be no well-defined “correct” output.

Interpretability: Many unsupervised learning algorithms, such as clustering methods, produce results that are not easily
interpretable. The meaning and significance of the clusters or patterns discovered may not be obvious, making it challenging to
draw meaningful insights.

Data Quality and Preprocessing: Unsupervised learning is highly sensitive to data quality. Noisy or incomplete data can lead to
misleading results. Data preprocessing and cleaning are often more critical in unsupervised learning compared to supervised
learning.

Reinforcement Learning
Understanding Reinforcement Learning

In reinforcement learning (RL), the machine interacts with an environment and learns to make a sequence of
decisions to maximize a cumulative reward signal. Unlike supervised learning, reinforcement learning doesn’t rely
on labeled data. Instead, the program learns through trial and error, receiving feedback in the form of rewards or
penalties for its actions.

Real-world examples of reinforcement learning include:

Gaming: RL algorithms have achieved remarkable success in mastering complex games like chess, Go, and video games. The
machine learns by playing against itself or other opponents, optimizing its strategies over time.

Robotics: Reinforcement learning is applied to robotic control tasks, like learning to grasp objects, walk, or fly drones. Robots
learn through physical interaction with the environment.

Autonomous Vehicles: RL plays a crucial role in training self-driving cars to make decisions in real-time, such as lane changing,
braking, and navigating complex road conditions.

Recommendation Systems: Reinforcement learning can be used to optimize recommendations by learning to suggest content
or products that maximize user engagement or revenue.

Healthcare: In healthcare, reinforcement learning can be used for personalized treatment plans, drug discovery, and optimizing
patient care.

Advantages and Limitations of Reinforcement Learning

Advantages:
Versatility: RL is versatile and can handle a wide range of tasks, from games to robotics to recommendation systems. It excels in
situations where explicit rules are challenging to define.

Adaptability: RL models can adapt to changing environments and learn from real-time interactions, making them suitable for
dynamic scenarios.

Complex Decision-Making: RL is great for problems involving complex, sequential decision-making where the consequences of one
action affect future decisions.

u d a c i t y. c o m
A B e g i n n e r ’ s G u i d e to M ac h i n e L e a r n i n g F u n da m e n ta l s

Limitations:
Sample Inefficiency: RL algorithms often require a large number of interactions with the environment to learn effectively. This can
be impractical or costly in real-world applications.

Safety Concerns: RL agents can pick up unsafe behaviors during training, posing risks in critical applications like autonomous
vehicles.

Exploration vs. Exploitation Trade-off: Striking the right balance between exploration (trying new actions to learn) and exploitation
(choosing known good actions) can be challenging.

Reward Engineering: Designing appropriate reward functions that accurately reflect the desired behavior can be difficult, and
poorly designed rewards can lead to unintended outcomes.

Machine Learning vs.

Traditional Programming
Machine learning and traditional programming represent
two distinct approaches to solving problems in the world of
computer science and software development. Here, we’ll
dive into the differences between each, along with
frameworks on when to use one over the other.

Traditional Programming :

Rule-Based: In traditional programming, developers write explicit rules and instructions for the computer to follow. These rules are
based on a deep understanding of the problem domain.

Deterministic: Traditional programs produce deterministic outputs. Given the same input, they will always produce the same
output.

Limited Adaptability: Traditional programs are rigid and don’t adapt to changing data patterns or unforeseen circumstances
without manual code modification.

Machine Learning :

Data-Driven: In machine learning, the algorithm learns from data rather than relying on explicitly programmed rules. It discovers
patterns and relationships within the data.

Probabilistic: Machine learning models make predictions based on probabilities. The same input may yield different outputs due to
inherent uncertainty in the models.

Adaptive: Machine learning models can adapt and improve their performance over time as they encounter more data, making them
suitable for dynamic and evolving scenarios.

udacit .com
y
A B e g i n n e r ’ s G u i d e to M ac h i n e L e a r n i n g F u n da m e n ta l s

When A Machine Learning Model Is Needed

Deciding whether or not a machine learning model is required for your specific project or traditional programming
will be sufficient comes down to a few key factors – mostly hinging on the nature of the problem and resources
available to you.

A Machine Learning Model Is Needed When:

The problem involves processing large and complex datasets where manual rule specification would be impractical or ineffective.

The problem requires making predictions or decisions based on historical data, and the patterns within the data are not easily
discernible through traditional methods.

You have access to sufficient labeled data for training and evaluation.

There’s a need for adaptability and the ability to improve over time.

A Machine Learning Model Might Not Be Needed When:

The problem is well understood, and a deterministic solution is achievable through rule-based approaches.

The problem has strict, unchanging rules and constraints that don’t depend on data patterns.

You have limited access to data or labeled data, making it challenging to train a machine learning model effectively.

How Machine Learning Works

Now that we have a better grip on what machine learning is, its history, and when we should use it, we’re going to
explore how machine learning actually works – along with how you should approach building your own machine
learning models.

Data Preparation and Preprocessing

Data preparation and preprocessing are the cornerstone of successful machine learning projects and are typically
among the first steps in the process. They involve cleaning, transforming, and organizing raw data to make it
suitable for training and evaluation.

Data preparation plays a crucial role in machine learning. It ensures that the data used for training is of high quality,
which, in turn, leads to accurate model results. During data preparation, features are engineered to make the model
perform better. It also helps the model adapt to new, unseen data, making it more practical for real-world use. Data
preparation helps lay a strong foundation for machine learning models, ensuring they can make reliable predictions
and decisions.

Key Steps in Data Preprocessing

Data preprocessing involves several key steps:
Data Cleaning: Removing or handling missing values, outliers, and errors. For example, in a dataset of patient records, handling
missing age values by ascribing them to the mean age.

Feature Engineering: Creating new features or transforming existing ones to capture relevant information. For instance, in a text
analysis project, converting text data into numerical features using techniques like TF-IDF (“Term Frequency-Inverse Document
Frequency”).

Data Scaling: Scaling features to a common range (between 0 and 1, etc.) to ensure that features with larger ranges don’t dominate
the learning process.

Encoding Categorical Data: Converting categorical variables into numerical form, often using techniques like “one-hot encoding” (a
representation method that converts categorical variables into a binary vector, where each category is represented by a unique
binary value, and all others are set to zero, which enables machine learning algorithms to work with categorical data). For instance,
transforming “red,” “green,” and “blue” categories into binary features.

udacity.com
A B e g i n n e r ’ s G u i d e to M ac h i n e L e a r n i n g F u n da m e n ta l s

Data Preprocessing, In Action

Let’s walk through a real-world example that brings these concepts to life –
predicting house prices based on features like square footage, number of
bedrooms, and neighborhood. In this case, data preprocessing might involve:
Data Cleaning: Handling missing values in square footage or bedrooms by replacing
them with median values.

Feature Engineering: Creating a new feature representing the ratio of square footage
to the number of bedrooms.

Data Scaling: Scaling all feature values to a common range to ensure they have equal
importance.

Building Machine Learning Models

T raining, T esting, and Model Evaluation
In the machine learning workflow, the training phase involves the model learning from the provided training data.
During this stage, the model adjusts its internal parameters through iterative processes to minimize prediction
errors, effectively capturing patterns and relationships within the data. Once the training is complete, the model’s
performance is assessed in the testing phase, where it encounters a separate dataset known as testing data. This
evaluation process allows us to gauge how effectively the model generalizes its learned knowledge to make
predictions on new, unexpected examples, providing valuable insights into its overall performance and reliability.

U nderstanding Model Evaluation Metrics

Model evaluation metrics are important tools for assessing the performance of machine learning models. These
metrics play a pivotal role in understanding how well a model is performing on a given task to help identify areas
for improvement. Typically, issues in models fall into one of two categories: classification (which involves assigning
data points to discrete categories), or regression (which deals with predicting continuous numerical values). Here
are some examples of which evaluation metrics to lean on for each.

For classification problems:

Accuracy: Accuracy is perhaps the most intuitive metric, measuring the proportion of correctly classified instances out of the total.
While it provides a general sense of model performance, it may not be suitable for imbalanced datasets where one class dominates
the others.

Precision: Precision focuses on the ratio of true positive predictions to all positive predictions made by the model. It helps in
identifying how well the model avoids false positives, making it particularly important in applications where false positives have
significant consequences, such as fraud detection.

Recall (Sensitivity): Recall, also known as sensitivity or true positive rate, measures the proportion of true positive predictions to all
actual positive instances. It provides insights into the model’s ability to identify all relevant instances of a particular class, which is
crucial in scenarios where missing positive cases is a concern, like disease diagnosis.

F1-Score: The F1-score is the mean of precision and recall, offering a balanced measure that considers both false positives and false
negatives. It’s valuable when you need to strike a balance between precision and recall, especially when there’s an uneven class
distribution.

For regression problems, evaluation metrics focus on quantifying the difference between predicted and
actual values:
Mean Absolute Error (MAE): MAE calculates the average absolute difference between the predicted and actual values. It provides a
straightforward measure of prediction accuracy and is less sensitive to outliers.

Mean Squared Error (MSE): MSE computes the average squared difference between predicted and actual values. It amplifies the
impact of larger errors, making it sensitive to outliers but still valuable for assessing model performance.

udacity.com
A B e g i n n e r ’ s G u i d e to M ac h i n e L e a r n i n g F u n da m e n ta l s

These evaluation metrics collectively offer a comprehensive view of a model’s strengths and weaknesses. By
analyzing these metrics, data scientists and machine learning practitioners can make informed decisions about
model selection, optimization, and deployment.

Machine Learning Basics FAQs

What is the difference between AI and machine learning?
AI (Artificial Intelligence) is a broad field of computer science focused on creating machines or systems that can
perform tasks that typically require human intelligence. AI encompasses various techniques, including machine
learning.

Machine learning, on the other hand, is a subset of AI. It involves training algorithms to learn from data and make
predictions or decisions without being explicitly programmed. In essence, machine learning is a methodology used
to achieve AI goals – so, while all machine learning is AI, not all AI is machine learning.

Are there 4 basic AI concepts?

Yes, there are four fundamental concepts in AI:
Machine Learning: This concept involves training algorithms to learn patterns and make predictions or decisions based on data.

Neural Networks: Neural networks are a type of model inspired by the structure of the human brain. They are used in deep
learning, a subfield of machine learning, to solve complex tasks like image recognition and natural language processing.

Natural Language Processing (NLP): NLP focuses on enabling machines to understand, interpret, and generate human language. It
has applications in chatbots, translation, and sentiment analysis.

Computer Vision: Computer vision is about enabling computers to interpret and understand visual information from the world,
such as images and videos. It plays a crucial role in areas like facial recognition and autonomous vehicles.

What should I learn first before machine learning?

Before diving into machine learning, it’s helpful to have a strong foundation in the following areas:
Programming: Familiarize yourself with a programming language like Python, as it’s widely used in the machine learning
community.

Mathematics: Brush up on essential mathematical concepts, especially linear algebra and calculus, which are fundamental to
understanding machine learning algorithms.

Statistics: Understand basic statistics, including concepts like mean, median, and standard deviation, as they play a crucial role in
data analysis and modeling.

Data Analysis: Learn how to work with data, including data cleaning, visualization, and exploratory data analysis.

Ready to jumpstart your machine learning journey?

There is so much to learn when it comes to machine learning, but truthfully, the space is closer to the starting line
than it is to the finish line! There’s room for innovators from all different walks of life and backgrounds to make their
mark on this industry of the future. Are you one of them? If so, we invite you to explore Udacity’s School of Artificial
Intelligence, and related Nanodegree programs. Our comprehensive curriculum and hands-on projects will equip
you with the skills and knowledge needed to excel in this rapidly growing field.

udacity.com
Learn More at
www.udacity.com

udacity.com