1. What is Machine Learning?
1.1 Agenda
Estimated reading time: ~10 minutes
Learning Outcomes
- Define Machine Learning with technical precision and distinguish it from AI broadly
- Describe the three main learning paradigms and when each applies
- Identify the specific problem types within supervised learning (classification vs. regression)
- Explain why data quality is more critical than algorithm choice
1.2 Glossary
| Term | Quick Explanation |
|---|---|
| Supervised Learning | Học có giám sát — thuật toán học từ dữ liệu đã được gán nhãn (labeled data): biết cả input lẫn expected output. |
| Unsupervised Learning | Học không giám sát — tìm pattern trong dữ liệu chưa có nhãn (unlabeled data), không biết output trước. |
| Reinforcement Learning | Học tăng cường — agent (tác nhân) học bằng cách thử nghiệm (trial and error) và nhận phần thưởng (reward) hoặc hình phạt (penalty). |
| Label | Nhãn — kết quả đúng (ground truth) được gán cho mỗi mẫu dữ liệu trong supervised learning. |
| Feature | Đặc trưng — biến đầu vào (input variable) dùng để dự đoán. Ví dụ: diện tích nhà, số phòng, vị trí → đều là features để dự đoán giá nhà. |
| Classification | Bài toán phân loại — dự đoán một nhãn rời rạc (discrete label) từ tập hữu hạn. Ví dụ: spam/not spam, cat/dog/car. |
| Regression | Bài toán hồi quy — dự đoán một giá trị liên tục (continuous value). Ví dụ: giá nhà, nhiệt độ ngày mai. |
| Clustering | Phân cụm — nhóm các điểm dữ liệu tương đồng với nhau mà không cần biết nhãn trước. |
| Anomaly Detection | Phát hiện bất thường — xác định điểm dữ liệu khác biệt đáng kể so với phần còn lại. |
2. Problem Statement
Traditional (truyền thống) software requires a developer to anticipate (lường trước) and encode (mã hóa) every possible decision path. This breaks down when:
- The rules are too complex to enumerate (liệt kê đầy đủ) — e.g., "what makes an email spam?"
- The environment changes — a fraud pattern from 2020 is different from 2025
- The data is unstructured (phi cấu trúc) — images, audio, free text cannot be processed by
if-thenlogic
Machine Learning solves this by inverting the paradigm (đảo ngược mô hình): instead of writing rules, you provide data + outcomes, and the algorithm learns the rules.
3. What is Machine Learning?
3.1 Definition
Machine Learning (ML) is a subset of Artificial Intelligence where computational systems learn to improve their performance on a specific task automatically through exposure (tiếp xúc) to data, without being explicitly (tường minh) programmed for every scenario.
3.2 Definition Anatomy
- "Subset of AI" — ML is one approach within the broader (rộng hơn) AI field. Rule-based systems are also AI but not ML.
- "Automatically" — The key word. The model's internal parameters (tham số nội tại) adjust themselves through training, without a human writing new rules.
- "Exposure to data" — ML systems require data to function. More quality data typically means better performance.
- "Without being explicitly programmed" — The developer writes the learning algorithm, not the decision logic. The decision logic emerges (xuất hiện) from the data.
3.3 The Core Inversion
4. Types of Machine Learning
4.1 Overview
AI-900 scope: Focuses primarily (chủ yếu) on Supervised and Unsupervised learning. Semi-supervised and Reinforcement Learning are introduced conceptually (theo khái niệm) only.
4.2 Supervised Learning
Definition: The algorithm learns from labeled data — each training example has both an input and a known correct output.
Two core problem types:
| Problem Type | Output | Example |
|---|---|---|
| Classification | Discrete label (nhãn rời rạc) from a finite set | Email → Spam / Not Spam |
| Regression | Continuous numeric (số liên tục) value | Square meters → House price |
Classification examples:
- Binary (nhị phân): Spam vs. not spam, fraud vs. legitimate
- Multi-class (đa lớp): Image → "cat", "dog", or "car"
- Multi-label (đa nhãn): News article → ["sports", "finance"] simultaneously (đồng thời)
Regression examples:
- House price prediction given area, location, number of rooms
- Expected (kỳ vọng) sales revenue next quarter
- Patient hospital stay duration (thời gian lưu viện) in days
4.3 Unsupervised Learning
Definition: The algorithm finds patterns in unlabeled data — there are no correct answers provided (được cung cấp). The model must discover (khám phá) structure on its own.
| Task | What It Does | Example |
|---|---|---|
| Clustering | Groups similar (tương đồng) data points | Customer segmentation (phân khúc) by purchasing behavior |
| Anomaly Detection | Flags data points that deviate (lệch) significantly from the norm (chuẩn) | Unusual transaction in banking |
| Dimensionality Reduction | Compresses (nén) high-dimensional data while preserving (giữ nguyên) information | Visualizing complex datasets in 2D |
4.4 Reinforcement Learning (Conceptual)
An agent (tác nhân) learns by interacting with an environment (môi trường):
- Takes an action (hành động)
- Receives a reward (phần thưởng) or penalty (hình phạt)
- Updates its strategy (chiến lược) to maximize (tối đa hóa) cumulative (tích lũy) reward
Examples: Game-playing AI (AlphaGo), robot navigation (điều hướng), recommendation system optimization.
AI-900 note: Reinforcement Learning is not directly tested. Understanding the concept at a high level is sufficient (đủ).
5. Choosing the Right Learning Type
6. Real-World Examples by Learning Type
| Scenario | Type | Why |
|---|---|---|
| Email spam detection | Supervised — Classification | Known labels: spam or not spam |
| House price prediction | Supervised — Regression | Output is a continuous number |
| Customer segmentation (phân khúc khách hàng) | Unsupervised — Clustering | No predefined groups; model discovers them |
| Credit card fraud detection | Unsupervised — Anomaly Detection | Fraudulent (gian lận) transactions are rare outliers (điểm ngoại lai) |
| Game-playing AI (Chess, Go) | Reinforcement Learning | Agent learns strategy through millions of games |
| Product recommendation | Can be Supervised or Unsupervised | Supervised if user ratings available; collaborative filtering (lọc cộng tác) if not |
7. Discussion Questions
Q1 — Labeling Cost: A company wants to build a sentiment classifier (bộ phân loại cảm xúc) for 1 million customer reviews. Manual labeling (gán nhãn thủ công) costs $0.10/review. Their budget is $5,000. What problem does this create, and what approaches could address the labeling bottleneck (điểm nghẽn gán nhãn) without collecting 1M labeled examples?
Q2 — Choosing the Type: A hospital wants to group patients by disease progression (tiến triển bệnh) patterns to tailor treatment plans — but they have no historical records linking patient profiles to specific outcomes. Which learning type is appropriate? What are the risks of applying the wrong type here?
Q3 — The Label Quality Problem: A bank trains a fraud detection model using historical transaction (giao dịch) records where the "fraud" label was determined by a human investigator (điều tra viên). If investigators only flagged (đánh dấu) 60% of actual fraud cases (the other 40% were missed), what does the model learn? How does this propagate (lan truyền) bias into production systems?
Made by Anh Tu - Share to be share