AI & ML Basics¶
1. The Big Picture: AI vs ML vs Deep Learning¶
These three terms are often used interchangeably, but they have a clear hierarchy:
Artificial Intelligence (AI)
└── Machine Learning (ML)
└── Deep Learning (DL)
Think of it like this: - AI is the broad goal: making machines behave intelligently. - ML is one approach to achieving AI: instead of programming rules, you let the machine learn from data. - Deep Learning is a specific ML technique inspired by the human brain, using layered neural networks.
2. Traditional Programming vs Machine Learning¶
Traditional Programming¶
You explicitly write all the rules. The computer follows them exactly.
Input + Rules → Output
Example: A spam filter where you manually write: “if the email contains the word ‘lottery’, mark as spam.”
Limitations: Brittle. You can’t anticipate every case. Rules become unmanageable at scale.
Machine Learning¶
You provide data and the expected outputs. The machine figures out the rules itself.
Input + Output (data) → Rules (model)
Example: You feed thousands of emails labeled “spam” or “not spam”. The model learns the patterns on its own.
Key insight: ML is about pattern recognition from experience, not explicit instructions.
Deep Learning¶
A subset of ML that uses neural networks with many layers (“deep” = many layers). Excels at complex, unstructured data like images, audio, and text.
Example: Recognizing faces in photos, transcribing speech, generating human-like text (like ChatGPT or Gemini).
Key insight: Deep Learning powers most modern AI breakthroughs, including Large Language Models (LLMs).
3. Types of Machine Learning¶
3.1 Supervised Learning¶
The model learns from labeled data — examples where the correct answer is already known.
- How it works: You provide input-output pairs. The model learns to map inputs to outputs.
- Goal: Predict the output for new, unseen inputs.
| Input | Label (Output) |
|---|---|
| Email text | Spam / Not Spam |
| House features | Price |
| Image of a cat | “Cat” |
Common tasks: - Classification — assign a category (spam or not, dog or cat) - Regression — predict a number (house price, temperature)
Real-world example: Google Photos recognizing faces, fraud detection in banking.
3.2 Unsupervised Learning¶
The model learns from unlabeled data — no correct answers are provided. It discovers hidden patterns or structure on its own.
- How it works: The model groups or organizes data based on similarity.
- Goal: Find structure where none was explicitly defined.
Common tasks: - Clustering — group similar items together (e.g., group customers by behavior) - Dimensionality Reduction — simplify complex data while preserving meaning - Anomaly Detection — identify unusual patterns (e.g., network intrusion detection)
Real-world example: Spotify grouping songs into playlists, customer segmentation in marketing.
3.3 Reinforcement Learning (RL)¶
The model (called an agent) learns by interacting with an environment and receiving rewards or penalties.
- How it works: The agent takes actions, receives feedback (reward/penalty), and adjusts its strategy over time.
- Goal: Maximize cumulative reward.
- Analogy: Training a dog with treats — good behavior gets rewarded.
Real-world examples: - Game-playing AI (AlphaGo, chess engines) - Robotics (teaching a robot to walk) - Ad bidding optimization - RLHF (Reinforcement Learning from Human Feedback) — used to align LLMs like Gemini with human preferences
Quick Comparison¶
| Supervised | Unsupervised | Reinforcement | |
|---|---|---|---|
| Data | Labeled | Unlabeled | No dataset needed |
| Feedback | Correct answers given | No feedback | Reward/penalty signal |
| Goal | Predict | Discover patterns | Maximize reward |
| Example | Email spam filter | Customer segmentation | Game-playing AI |
4. The Machine Learning Lifecycle¶
Building an ML model is not a one-step process. It follows a lifecycle with distinct stages:
1. Problem Definition
↓
2. Data Collection
↓
3. Data Preparation
↓
4. Model Training
↓
5. Model Evaluation
↓
6. Model Deployment
↓
7. Monitoring & Maintenance
↓
(loop back as needed)
Stage 1 — Problem Definition¶
Clearly define what you want to predict or automate. Is it a classification problem? Regression? What data do you have? What does “success” look like?
Stage 2 — Data Collection¶
Gather raw data from relevant sources: databases, APIs, sensors, user interactions, etc. More data generally means better models — but quality matters more than quantity.
Stage 3 — Data Preparation¶
This is typically the most time-consuming stage (~70-80% of the effort). - Cleaning: Remove duplicates, fix errors, handle missing values - Transformation: Normalize/scale values, encode categorical variables - Splitting: Divide data into training set, validation set, and test set - Feature engineering: Create new meaningful inputs from raw data
Stage 4 — Model Training¶
Feed the prepared data into a chosen algorithm. The model adjusts its internal parameters to minimize prediction errors. This is where “learning” actually happens.
Key concepts: - Algorithm selection: Which model type to use (decision tree, neural network, etc.) - Hyperparameters: Settings that control the training process (learning rate, number of layers) - Overfitting: When the model memorizes training data but fails on new data - Underfitting: When the model is too simple to capture patterns
Stage 5 — Model Evaluation¶
Test the model on data it has never seen before (the test set). Measure performance using metrics: - Accuracy — % of correct predictions - Precision / Recall — important for imbalanced datasets (e.g., fraud detection) - F1 Score — balance between precision and recall - RMSE — for regression tasks
If performance is insufficient, loop back to earlier stages.
Stage 6 — Model Deployment¶
Package the model and expose it via an API or integration so applications can use it in production. On Google Cloud, this typically happens through Vertex AI.
Stage 7 — Monitoring & Maintenance¶
ML models can degrade over time as real-world data changes — this is called model drift or data drift. Continuous monitoring ensures the model remains accurate and relevant.
5. Key Vocabulary Cheat Sheet¶
| Term | Definition |
|---|---|
| Feature | An input variable used to make a prediction (e.g., age, email text) |
| Label | The correct output/answer during training (e.g., “spam”) |
| Model | The trained artifact that makes predictions |
| Training | The process of a model learning from data |
| Inference | Using a trained model to make predictions on new data |
| Overfitting | Model too tailored to training data, fails on new data |
| Underfitting | Model too simple, misses patterns |
| Neural Network | A layered structure of nodes inspired by the human brain |
| Parameter | Internal values the model adjusts during training (e.g., weights) |
| Hyperparameter | Settings you choose before training (e.g., learning rate) |
| Data Drift | When real-world data changes and the model becomes less accurate |
| RLHF | Reinforcement Learning from Human Feedback — used to align LLMs |
6. How This Connects to Generative AI¶
Generative AI (like Gemini) is built on top of these foundations:
- It uses Deep Learning (specifically Transformer-based neural networks)
- It is trained with a mix of supervised learning (next-token prediction) and reinforcement learning (RLHF — human raters give feedback to improve quality and safety)
- The ML lifecycle applies to LLMs too: data collection → pre-training → fine-tuning → evaluation → deployment → monitoring
Understanding these basics is what allows you to reason about why GenAI models behave the way they do — and how to improve or correct their outputs.