A4.1.1 · Machine learning fundamentals
Types of Machine Learning
Every machine learning system falls into one of a small number of families, and which family it belongs to depends on one simple question: what does the algorithm actually get told about the "right answer" while it's learning?
What principles and approaches should be considered to ensure machine learning models produce accurate results ethically? Every statement in this topic comes back to this question, and knowing what a machine learning model actually is, and which of the five families it belongs to, is the first step toward being able to evaluate it responsibly.
Explainer
Machine learning and artificial intelligence get used almost interchangeably in everyday conversation, but they aren't the same thing.
Artificial intelligence (AI) is the broad goal of building systems that can perform tasks normally requiring human-level judgement. Machine learning (ML) is one specific way of reaching that goal: instead of a programmer writing out every rule by hand, the system is given data and left to work out its own rules for the task, adjusting itself as it sees more examples.
Worked example
Five machine learning families
What separates one type from another is how much the algorithm is told about the correct answer while it trains, and where that information (or lack of it) actually comes from. Click each one to see a distinct real-world scenario for it.
Reference
Six required real-world applications
The syllabus names six specific applications you're expected to know, each one naturally suited to a particular type (or combination of types) of machine learning.
| Application | Best-fit type(s) | Why |
|---|---|---|
| Market basket analysis | Unsupervised | No "correct" combination of products is ever labelled in advance, the algorithm just looks for items that keep turning up together. |
| Medical imaging diagnostics | Supervised, deep learning | Needs scans already labelled with a confirmed diagnosis, and a many-layered network able to pick out subtle visual detail a human might miss. |
| Natural language processing | Deep learning, transfer learning | Language structure is intricate and layered; a model pre-trained on huge amounts of text gets fine-tuned for the specific task at hand. |
| Object detection and classification | Supervised, deep learning | Needs images already labelled with what's inside them, and a many-layered network to recognise shapes built out of smaller shapes. |
| Robotics navigation | Reinforcement | There's no labelled "correct path" to learn from in advance, so the robot learns by trying routes and being rewarded for avoiding obstacles. |
| Sentiment analysis | Supervised | Needs text already labelled positive or negative, so the model can learn the mapping from wording to sentiment. |
Check your understanding
A supermarket wants to know which products are frequently bought together, without knowing in advance what those combinations might be. Which type of machine learning best fits this problem?
Common mistake
RecurringA genuinely documented pattern: candidates describe deep learning as if it were a sixth, entirely separate category sitting alongside supervised, unsupervised, reinforcement and transfer learning. It isn't. Deep learning describes HOW an algorithm is built, using a many-layered neural network, rather than WHAT it's being asked to learn. A supervised learning task, an unsupervised one, or a reinforcement one can each be built using deep learning or without it.
Treats deep learning as a sixth, separate category: "There are five machine learning categories: supervised, unsupervised, reinforcement, transfer, and deep learning." Lists deep learning alongside the other four as if it answered the same question they do, when it actually answers a different one entirely.
Places deep learning on its own dimension: "Supervised, unsupervised, reinforcement and transfer learning describe what an algorithm learns from. Deep learning describes how it's built, a many-layered neural network, and any of those four approaches can be implemented using it." Correctly separates the "what it learns from" question from the "how it's built" question.
Revision questions
Past-paper style. Write an answer, then reveal the marking points and mark yourself.
[2]Define the term machine learning, and state how it differs from artificial intelligence.
Marking points
评分要点
- Machine learning is an approach to building intelligent systems where the system learns patterns from data rather than following explicitly programmed rules.
- It is a subfield of artificial intelligence: AI is the broad goal of intelligent behaviour, ML is one method of reaching it, defined by learning from data.
- 机器学习是构建智能系统的一种方法,系统从数据中学习模式,而不是遵循明确编程的规则。
- 它是人工智能的一个子领域:人工智能是智能行为这一广阔目标,机器学习是达成它的一种方法,以从数据中学习为特征。
[4]Distinguish between supervised learning and unsupervised learning, referring to what each does and does not receive as part of its training data.
Marking points
评分要点
- Supervised learning trains on labelled data: every example includes the correct output, and the model learns to predict that output for new inputs.
- Unsupervised learning trains on unlabelled data: there are no target outputs, and the model finds structure such as clusters or reduced dimensions on its own.
- So the key difference is whether the training data includes the correct answers.
- 监督学习在带标签数据上训练:每个样例都包含正确输出,模型学会为新输入预测该输出。
- 无监督学习在无标签数据上训练:没有目标输出,模型自行发现结构,如聚类或降维。
- 所以关键区别在于训练数据是否包含正确答案。
[6]A city council wants to build a system that adjusts traffic light timing in real time based on the outcome of its previous decisions. Identify the most suitable type of machine learning for this problem and justify your choice.
Marking points
评分要点
- The most suitable type is reinforcement learning.
- The system makes a decision (a light timing), observes the outcome (traffic flow) and receives a reward or penalty based on that outcome.
- Over many cycles it adjusts its policy to maximise the long-run reward.
- Supervised learning is a poor fit because there is no labelled data set of correct light timings for every traffic state.
- Unsupervised learning is a poor fit because the task is to choose actions, not to find structure in data.
- Reinforcement learning matches because decisions are judged by their downstream outcomes over time, which is exactly what a reward signal captures.
- 最合适的类型是强化学习。
- 系统做出一个决策(一个信号灯配时),观察结果(交通流量),并根据该结果获得奖励或惩罚。
- 在多个周期中,它调整策略以最大化长期奖励。
- 监督学习不合适,因为不存在针对每种交通状态的正确配时的带标签数据集。
- 无监督学习不合适,因为任务是选择动作,而不是在数据中寻找结构。
- 强化学习匹配,因为决策是由其随时间产生的后续结果来评判的,而这正是奖励信号所捕获的。
Flashcards
Active recall for this statement’s key terms.
Keep this statement’s practice with you
Free account: unlock every revision question and flashcard for A4.1.1, and every other IB CS 2027 statement.