AI and ML Fundamentals
Learn the key concepts and terminology of artificial intelligence and machine learning to effectively document these technologies.
Table of Contents
Now that weâve packed our bags with curiosity, letâs step into the world of Artificial Intelligence (AI) and Machine Learning (ML).
You donât need a PhD in math to understand these concepts. Just bring your imagination, a few metaphors, and maybe a cup of coffee.
AI is the big umbrella, ML is the clever part inside, and deep learning is its overachieving cousin.
What Is AI? What Is ML? And Why Should You Care?
Letâs break it down as simply as possible.
Artificial Intelligence (AI)
Think of AI as the big idea: teaching machines to behave intelligentlyâlike playing chess, recognizing your face in a photo, or making movie recommendations.
Machine Learning (ML)
ML is how machines learn. Instead of being programmed with strict rules, they look at data and learn from patterns. Like how you get better at riding a bicycle by practicing.
Deep Learning
Deep learning is a subset of ML. It uses neural networks with many layers to learn complex thingsâlike understanding spoken language or recognizing animals in photos.
Types of Machine Learning: How Machines Learn
Supervised Learning
This is like learning with a teacher. You show the algorithm examples with answers, and it learns to make predictions.
Example: Teach it the difference between cats and dogs using labeled images.
Common algorithms:
- Linear and Logistic Regression
- Support Vector Machines
- Decision Trees and Random Forests
- Neural Networks
Unsupervised Learning
Here, the machine learns without labeled data. It finds patterns on its own.
Example: Grouping customers based on purchasing habits without telling the algorithm what the groups should be.
Common algorithms:
- K-means Clustering
- Hierarchical Clustering
- Principal Component Analysis
- Association Rules
Reinforcement Learning
The algorithm learns by doing, receiving rewards or penalties. Itâs trial and error at scale.
Example: Learning to play a video game by winning or losing points based on actions.
Common algorithms:
- Q-Learning
- Deep Q Networks (DQN)
- Policy Gradient Methods
- Actor-Critic Methods
The Machine Learning Pipeline
Think of this as the recipe for building an ML solution.
1. Data Collection and Preparation
Start with gathering data. It might come from:
- Databases
- Sensors or devices
- User interactions
- Web scraping
- Public datasets
Then clean the data:
- Handle missing values
- Remove duplicates
- Fix incorrect data types
- Normalize values
2. Exploratory Data Analysis (EDA)
Understand your data using:
- Statistical summaries
- Visualizations
- Correlation checks
- Outlier detection
3. Feature Engineering
Features are what the model uses to make decisions:
- Create new features
- Scale or normalize data
- Encode text or categories
- Select the most relevant features
4. Model Training and Selection
Pick the right algorithm for the job, then:
- Split data into training and validation sets
- Train the model
- Tune parameters (hyperparameters)
- Use cross-validation to test stability
5. Evaluation
Use metrics to see how well the model works.
How do we know if our model is any good? Let's talk metrics!
| Task Type | Metric | In Plain English | When to Use It | Score Range |
|---|---|---|---|---|
| Classification | Accuracy đŻ | "How often is my model right?" | When all mistakes cost the same |
0-100% (higher is better)
|
| Precision đ | "When it predicts 'yes', how often is it correct?" | When false alarms are expensive |
0-100% (higher is better)
|
|
| Recall đ¸ď¸ | "What percentage of actual 'yes' cases did it catch?" | When missing positives is costly |
0-100% (higher is better)
|
|
| F1 Score âď¸ | "The harmony between precision and recall" | When you need balance between both |
0-100% (higher is better)
|
|
| Regression | MAE đ | "On average, how far off are my predictions?" | When all errors matter equally | Lower is better |
| MSE đ | "Like MAE, but big mistakes are punished more" | When large errors are particularly bad | Lower is better | |
| RMSE đ | "MSE, but brought back to original units" | When you want errors in original units | Lower is better | |
| R² đ | "What percentage of the variation does my model explain?" | When comparing models across datasets |
0-1 (higher is better)
|
Pro tip: No single metric tells the whole story. Always consider multiple metrics and the real-world impact of your model's predictions!
6. Deployment
Put the model into production:
- Expose it via an API
- Integrate it into apps
- Set up monitoring
7. Monitoring and Maintenance
Even after deployment, your model needs care:
- Watch performance
- Retrain as new data arrives
- Handle concept drift
Common Machine Learning Pitfalls
Overfitting
The model memorizes the training data too well and fails on new data.
Fix it by:
- Simplifying the model
- Adding more data
- Using regularization
Underfitting
The model is too simple and fails to learn from the data.
Fix it by:
- Using a more complex model
- Adding more useful features
- Allowing more training time
Data Leakage
The model accidentally uses information it shouldnât have during training.
Fix it by:
- Carefully splitting data
- Avoiding future info
- Designing features responsibly
Class Imbalance
When one class dominates, the model might ignore the smaller class.
Fix it by:
- Resampling techniques
- Generating synthetic data
- Using appropriate evaluation metrics
Comparing Algorithms
Choosing the right algorithm is like picking the right tool for a job. Let's meet the contenders!
| Algorithm | What It's Good At | Where It Struggles | Perfect For | Complexity |
|---|---|---|---|---|
| Linear Regression Regression |
|
|
|
Beginner Friendly |
| Logistic Regression Classification |
|
|
|
Beginner Friendly |
| Decision Trees Both |
|
|
|
Intermediate |
| Random Forest Both |
|
|
|
Intermediate |
| Support Vector Machine Both |
|
|
|
Advanced |
| Neural Networks Both |
|
|
|
Expert Level |
đĄ Pro tip: Simpler algorithms often outperform complex ones on smaller datasets. Don't reach for a neural network when linear regression will do!
Neural Networks Explained
Neural networks are inspired by the human brain, but they work on numbers.
Structure
- Input Layer: Takes in raw data (like pixels or numbers)
- Hidden Layers: Process information and find patterns
- Output Layer: Gives the final prediction
- Neurons: Each node that does some math
- Weights: Control how strongly inputs influence outputs
How Training Works
- Data goes forward through the layers (forward pass)
- Compare prediction to actual result (loss)
- Send feedback backward to update weights (backpropagation)
- Repeat until it gets better
Common Types
- Convolutional Neural Networks (CNNs): Best for images
- Recurrent Neural Networks (RNNs): Great for sequences and time
- Transformers: Power modern language models
- Generative Adversarial Networks (GANs): Create new content
The Neural Network Family: Each One Has Its Superpower
| Network Type | How It Works | What It's Great At | Famous Examples | Difficulty Level |
|---|---|---|---|---|
|
CNN
Convolutional Neural Networks
The Visual Specialists
|
Uses filters that slide across images to detect patterns like edges, textures, and shapes at different levels of abstraction. |
|
ResNet
VGG
Inception
YOLO
|
Moderate |
|
RNN
Recurrent Neural Networks
The Memory Keepers
|
Processes sequences by maintaining a memory of what came before, making them perfect for data where order matters. |
|
LSTM
GRU
Bi-LSTM
|
Challenging |
|
TF
Transformers
The Attention Masters
|
Uses attention mechanisms to weigh the importance of different parts of input data, allowing them to process entire sequences at once. |
|
BERT
GPT
T5
DALL-E
|
Advanced |
|
GAN
Generative Adversarial Networks
The Creative Duos
|
Two networks compete: one creates content, the other judges its authenticity, driving each other to improve through competition. |
|
StyleGAN
CycleGAN
Pix2Pix
|
Advanced |
đĄ Did you know? The largest neural networks today (like GPT-4) have hundreds of billions of parameters - that's like having more connections than there are stars in our galaxy!
The Limits of AI
Data Dependency
Garbage in, garbage out. Poor data = poor model.
Lack of Understanding
AI doesnât actually âknowâ thingsâit just recognizes patterns.
Black Box Models
Deep models are hard to explain, even to experts.
Fragile in New Situations
Trained for summer? It might fail in winter unless retrained.
High Resource Demand
Training big models requires massive computing power.
Ethical Considerations in AI
AI isnât just technicalâitâs deeply human.
Bias and Fairness
Biased training data leads to biased outcomes.
Privacy
Machine learning often needs personal data. This raises red flags.
Transparency
People should understand decisions that affect them.
Accountability
When AI fails, who takes the blame?
Environmental Impact
Training large models contributes to carbon emissions.
What This Means for Documentation
Understanding these fundamentals helps you:
Provide Technical Depth
- Create layered content
- Use analogies and visuals
Communicate Limitations Clearly
- Be transparent about edge cases
- Manage user expectations
Track Change and Versioning
- Show how models evolve
- Explain whatâs different between versions
Write Ethically
- Mention fairness, privacy, and data origin
- Include known risks and how they are handled
Exercise: Identify the ML Approach
Pick an application and answer:
- What type of learning is used?
- Is it classification, regression, clustering, etc.?
- What kind of data does it need?
- What might be hard to explain in the documentation?
Examples:
- Spam filter
- Product recommendation
- Fraud detection
- Customer segmentation
- Self-driving car
- Stock price prediction
Want to Learn More?
Books
- âThe Hundred-Page Machine Learning Bookâ by Andriy Burkov
- âInterpretable Machine Learningâ by Christoph Molnar
- âArtificial Intelligence: A Guide for Thinking Humansâ by Melanie Mitchell
Free Courses
Helpful Resources
Whatâs Coming Up Next?
Next, weâll explore how to explain AI and ML systems to different types of readersâdevelopers, decision-makers, and everyday users.
Youâll learn how to adjust your writing style, use the right tone, and build clarity even for the most complex concepts.
Letâs move from understanding AI to helping others understand it.
Frequently Asked Questions About AI-ML Fundamentals
Get answers to common questions about AI and ML concepts, terminology, and documentation approaches.
AI and ML Basics
Artificial Intelligence (AI) is the broader concept of machines being able to carry out tasks in a way that we would consider âsmartâ or âintelligent.â Machine Learning (ML) is a subset of AI where algorithms learn patterns from data without being explicitly programmed for specific tasks. In documentation, itâs important to use these terms preciselyâML refers specifically to systems that learn from data, while AI encompasses a wider range of approaches including rule-based systems, expert systems, and machine learning.
The main types of machine learning are: 1) Supervised Learning, where models learn from labeled data to make predictions; 2) Unsupervised Learning, where models find patterns in unlabeled data; 3) Reinforcement Learning, where agents learn through interaction with an environment and feedback; and 4) Semi-supervised Learning, which uses a combination of labeled and unlabeled data. When documenting ML systems, specifying the learning approach helps users understand the data requirements and potential limitations.
Neural networks are computational models inspired by the human brainâs structure, consisting of interconnected âneuronsâ that process information. Deep learning specifically refers to neural networks with multiple layers (deep neural networks). While all deep learning uses neural networks, not all neural networks are deep. In documentation, itâs helpful to specify whether a system uses simple neural networks or deep learning, as this impacts computational requirements, training time, and the amount of data needed.
Practical Documentation
AI systems differ from traditional software because they learn from data rather than following explicit programming instructions. This means documentation must cover data sources, training processes, model limitations, and probabilistic performance metricsâaspects not typically found in traditional software docs. Additionally, AI documentation must address potential biases, explain how the system behaves in edge cases, and provide guidance for handling model drift over time.
Effective AI documentation should layer information for different audiences. Technical details (model architecture, hyperparameters, training methods) should be accessible to data scientists and engineers, while higher-level information (capabilities, limitations, use cases) should be presented in accessible language for business users and decision-makers. Each layer should contain appropriate detail without overwhelming users. Documentation should also cross-reference between levels, allowing users to dive deeper into technical aspects when needed.
AI model documentation should include relevant metrics based on the task type: classification (accuracy, precision, recall, F1-score), regression (RMSE, MAE, R²), ranking (NDCG, MAP), generative tasks (perplexity, BLEU score), and reinforcement learning (cumulative reward). It should also document metrics across different data slices to identify potential biases or performance gaps. Additionally, include real-world performance indicators that matter to end-users, and be transparent about evaluation methodology and test data characteristics.
AI and ML Fundamentals Quiz
What is the relationship between Artificial Intelligence (AI) and Machine Learning (ML)?
Found value in the course?
Your support fuels my work!
Have an issue?
Help improve the course by reporting an issue
