2. AI, Machine Learning, and Deep Learning
2.1 Agenda
Estimated reading time: ~10 minutes
Learning Outcomes
- Describe the containment (bao hàm) hierarchy: AI ⊃ ML ⊃ DL
- Explain why each layer exists and what problem it solves
- Select the most appropriate (phù hợp) layer for a given problem
2.2 Glossary
| Term | Quick Explanation |
|---|---|
| Algorithm | Quy trình từng bước để máy học pattern từ dữ liệu. |
| Neural Network | Mô hình tính toán gồm các lớp node kết nối — cơ sở của Deep Learning. |
| Feature Engineering | Bước tiền xử lý thủ công: con người tự chọn đặc trưng (features) quan trọng từ dữ liệu thô. |
| Representation Learning | Khả năng tự học đặc trưng từ dữ liệu — đặc điểm nổi bật của Deep Learning. |
3. Problem Statement
A single label — "Artificial Intelligence" — is applied to systems ranging from a 1990s chess engine to GPT-4. This ambiguity (sự mơ hồ) causes practitioners (người thực hành) to misdiagnose (chẩn đoán sai) requirements — proposing deep learning when classical ML would achieve 95% of the performance at 10% of the cost.
4. The Containment Hierarchy
Each inner ring is a specialization (chuyên biệt hóa) of the outer — not a replacement.
5. Artificial Intelligence (Outer Ring)
AI is the broadest category — any technique that enables a machine to mimic (bắt chước) cognitive (nhận thức) functions, including rule-based systems and learning-based systems.
Systems in the AI ring but outside ML: expert systems, Minimax chess engines, Dijkstra's pathfinding (tìm đường) — all encode (mã hóa) logic manually, with no learning from data.
6. Machine Learning (Middle Ring)
ML is a subset of AI where systems improve performance through exposure (tiếp xúc) to data, without being explicitly (tường minh) programmed for every scenario.
Core limitation: Classical ML requires Feature Engineering — the manual process of deciding which aspects of raw data are informative (có giá trị thông tin). This bottleneck (điểm nghẽn cổ chai) is what Deep Learning addresses.
7. Deep Learning (Inner Ring)
Deep Learning is a subset of ML using multi-layered (nhiều tầng) neural networks to learn hierarchical (phân cấp) representations directly from raw data — bypassing (bỏ qua) Feature Engineering.
For an image classification task, layers learn progressively (dần dần) abstract features:
7.1 Classical ML vs. Deep Learning
| Dimension | Classical ML | Deep Learning |
|---|---|---|
| Feature engineering | Manual (thủ công) | Automatic (tự động) |
| Data need | Moderate (vừa phải) | Large — millions of samples |
| Compute cost | Low | High — requires GPUs |
| Interpretability (khả năng giải thích) | Transparent | Opaque (mờ đục — black box) |
| Best for | Structured tabular data (dữ liệu bảng có cấu trúc) | Unstructured data (phi cấu trúc — ảnh, text, audio) |
8. Practical Decision Framework
9. Discussion Questions
Q1 — Classification Challenge: A startup claims "AI" matches candidates to roles via a scoring formula with manually set weights. Is this AI? ML? DL? Justify with definitions.
Q2 — The Data Constraint: A hospital builds a rare-disease diagnostic (chẩn đoán) tool with only 500 confirmed (được xác nhận) cases. Which layer is appropriate? What technique compensates (bù đắp) for the data constraint? (Hint: transfer learning.)
Q3 — Interpretability Trade-off: A bank must explain every loan rejection (từ chối khoản vay) by regulation. DL achieves 92% accuracy; a Decision Tree achieves 85%. Which do you deploy and why?
Made by Anh Tu - Share to be share