1. What is Machine Learning?
Machine Learning (ML) is a branch of Artificial Intelligence in which computers learn from data and experience rather than being given explicit instructions for every situation.
Traditional programming works like:
Input → Rules/Instructions → Output
Machine learning works more like:
Data → Algorithm → Patterns/Learning → Prediction/Decision
The machine examines data, identifies patterns, receives feedback, and adjusts its model so that it can improve over time.
A simple analogy from the course is learning to assemble IKEA furniture: you try something, see whether it works, learn from the feedback, and use that experience when assembling the next piece.
2. Why do we need Machine Learning?
Some problems cannot easily be solved by writing explicit rules.
For example, identifying spam emails is difficult with simple instructions because spammers can change words, use images, or disguise certain terms.
Instead, ML can examine thousands of emails, learn characteristics associated with spam, and then predict whether a new email is spam.
The basic process is:
Training data → Build model → Test model → Make predictions
An important concept is keeping training data separate from test data.
3. Machine Learning and Data
Machine learning is essentially about finding patterns in large amounts of data.
Examples mentioned in the course include:
- Search engines
- Weather applications
- Netflix recommendations
- YouTube recommendations
- Amazon product recommendations
- Automatic translation
- Speech-to-text
- Personalized social-media feeds
A particularly important point is that machines can sometimes discover patterns that humans may not notice.
Therefore, the quality and diversity of data are extremely important.
Poor data → Poor learning → Poor predictions
4. Four Ways Machines Learn
The course introduces four important approaches.
| Type | How the machine learns | Simple example |
|---|---|---|
| Supervised Learning | Learns from labelled examples | Spam vs. non-spam |
| Unsupervised Learning | Finds patterns in unlabelled data | Customer groups |
| Semi-supervised Learning | Uses some labelled + lots of unlabelled data | Partial classification |
| Reinforcement Learning | Learns through actions and rewards | Playing a game |
These distinctions are also consistent with current standard descriptions of ML learning paradigms.
Supervised Learning
Think of a teacher.
The machine receives examples where the correct answer is already known.
For example:
Email → Spam
Email → Not Spam
The machine learns the relationship and then classifies new emails.
Two common applications are:
- Classification — predicting categories
- Regression — predicting numerical values
Unsupervised Learning
There is no teacher providing the correct answer.
The machine looks at unlabelled data and tries to discover patterns or groups on its own.
For example, it might discover different groups of customers based on their behaviour.
Semi-supervised Learning
This combines the two.
You provide some labelled data, while the machine learns from a much larger quantity of unlabelled data.
Reinforcement Learning
This is essentially learning by doing.
The machine takes an action, receives a reward or penalty, and gradually learns which actions produce better results.
The course uses games such as Pong, Space Invaders and Go to explain this concept.
5. Q-Learning
Q-learning is introduced as an important reinforcement-learning technique.
The machine considers:
State → Action → Result/Reward
It gradually learns which actions produce better outcomes.
The course uses AlphaGo and AlphaGo Zero as examples of how reinforcement learning can allow machines to improve through enormous numbers of simulations.
6. Popular Machine Learning Algorithms
The course introduces several algorithms:
- Decision Trees
- K-Nearest Neighbour (KNN)
- K-Means Clustering
- Regression
- Bayesian algorithms / Naive Bayes
Very simple distinction
Decision Tree:
Makes decisions through a sequence of questions.
KNN:
Looks at nearby/similar examples to determine the likely classification.
K-Means:
Groups similar observations into clusters.
Regression:
Studies relationships between variables and can be used to predict numerical outcomes.
Naive Bayes:
Uses probability to estimate the likelihood of different outcomes, particularly for classification.
The source specifically notes that Naive Bayes assumes its predictors are independent of one another.
7. Bias and Variance
An important part of applying ML algorithms is understanding bias and variance.
They represent different types of prediction problems.
The goal isn't simply to eliminate one completely. Instead, the model needs to achieve an appropriate balance so that it performs well on new, unfamiliar data.
8. Three Major Challenges
The course identifies three practical challenges:
1. Asking the right questions
Having sophisticated ML technology isn't enough.
The business must first identify interesting and useful questions that machine learning can actually help answer.
The course gives a memorable analogy: having a Formula One race car but using it only to parallel park.
2. Separating training and testing data
You should not mix your training data with your test data.
Otherwise, the model may appear highly accurate because it has effectively already seen the answers.
The objective is generalizability — the model should perform well on unfamiliar data.
3. Don't become obsessed with selecting the “perfect” algorithm
Different algorithms have different strengths.
The course suggests that practitioners should focus on solving the business problem rather than spending excessive time searching for a theoretically perfect algorithm.
The entire course in one picture
Artificial Intelligence
↓
Machine Learning
↓
Machine learns from data instead of receiving every instruction explicitly
↓
Data → Algorithm → Pattern → Model → Prediction/Decision
And the four major learning approaches are:
Supervised → Learn from labelled examples
Unsupervised → Discover patterns
Semi-supervised → Some labelled + lots of unlabelled data
Reinforcement → Learn through actions and rewards
The 5 things I would remember for an exam/interview
- Machine Learning = learning patterns from data rather than relying entirely on explicit instructions.
- Training data is used to build/refine the model; test data evaluates it on unfamiliar data.
- Supervised = labelled data; Unsupervised = unlabelled data; Reinforcement = rewards/penalties.
- Common algorithms include Decision Trees, KNN, K-Means, Regression and Naive Bayes.
- Successful ML is not just about algorithms — good data, good questions and proper testing are equally important.
For someone coming from a finance/FP&A background, the most useful next step would be to translate these concepts into finance examples—forecasting, fraud detection, credit risk, customer segmentation, anomaly detection and financial reporting. That would make the course much easier to remember and apply.
No comments:
Post a Comment