Based on the AIF-C01 exam guideReport an error
data + rules → results, and ML is data + results → rules. ML fits when the rules are complex and there is enough data.Review is not memorization but a time to redraw how the concepts connect.
[Day1] The AI > ML > DL relationship
↓
[Day2] Learning types in ML (supervised, unsupervised, reinforcement)
↓
[Day3] Problems where ML fits vs. problems where it doesn't
↓
[Day4] Key terms (model, training, inference, feature, label, overfitting)
| Learning type | Answer (label) | Representative tasks |
|---|---|---|
| Supervised learning | Present | Classification (category), regression (number) |
| Unsupervised learning | Absent | Clustering, dimensionality reduction |
| Reinforcement learning | Replaced by a reward | Games, robots, control |
Let us sort out the two pairs that are often confused once more.
data + rules → resultsdata + results → rules| ML is a good fit | ML is a poor fit |
|---|---|
| The rules are complex or hard to define | The rules are simple and clear |
| There is enough data | There is almost no data |
| A small margin of error is acceptable | 100% accuracy and a complete explanation are mandatory |
Features + labels [training data]
↓ Training
[Model] completed
↓ Inference
New input (features) → predicted label
↓
Evaluate with [test data] → check for overfitting/underfitting
This week's content is the foundation of the "AI/ML fundamentals" domain of the AIF-C01 exam. Only if you understand this flow clearly can you follow later advanced topics such as generative AI and AWS AI services without strain.
| Trap | Correct understanding |
|---|---|
| "Deep learning is the biggest concept" | AI is the biggest. Deep learning is the smallest |
| "Classification and regression are the same" | Classification = category, regression = number |
| "Classification and clustering are the same" | They differ in whether answer labels exist |
| "Machine learning is always better" | For simple problems, the traditional approach is better |
| "It's fine to evaluate with the training data" | You must evaluate with test data to be honest |
| "Overfitting = insufficient training" | Overfitting is rather a state of having memorized too much |
The four days of content scatter if memorized separately, but when strung together in the order of questions, they become a single procedure for telling cases apart.
① Who created the rules?
├─ A person wrote the conditions directly ────► Rule-based (may be AI, but not ML)
└─ Learned from data ─────────────────────────► Machine learning (go to ②)
└ If many layers of neural networks, deep learning
② What did it learn from?
├─ Data with answers (labels) attached ───────► Supervised learning (go to ③)
├─ Data without answers ──────────────────────► Unsupervised learning (clustering, dimensionality reduction, anomaly detection)
└─ Rewards received as a result of actions ───► Reinforcement learning
③ What shape is the answer to be predicted?
├─ One of a fixed set of kinds ───────────────► Classification
└─ A number whose magnitude has meaning ──────► Regression
④ When performance comes up, look at the two scores
├─ Training is low and test is low ───────────► Underfitting
├─ Only training is high and test drops sharply ► Overfitting
└─ Both are high and the gap is small ────────► Generalizes well
An online bookstore says, "We want to find out in advance which orders will be returned." This one sentence touches everything from this week.
| Step | What happens at the bookstore | Concept involved |
|---|---|---|
| Judging necessity | A policy like "no returns under 30,000 KRW" is fine as code. But "which orders will be returned" cannot be written as rules, and there are hundreds of thousands of past order records | Day 3: conditions where ML fits |
| Technology level | It is tabular data such as order amount, delivery region, and purchase history. Deep learning is not necessarily needed | Day 1: distinguishing ML from deep learning |
| Learning type | Each past order has a "returned / not returned" record → supervised learning, and the answer is two categories → classification | Day 2: whether labels exist, category/number |
| Data composition | Order amount, region, and history are the features, and whether it was returned is the label | Day 4: features and labels |
| Building the model | Train on part of the records and score on the test data set aside. If training is 98% and test is 71%, it is overfitting | Day 4: data splitting, overfitting |
| Service | To learn the return likelihood immediately whenever an order comes in, use real-time inference; to check the day's orders in bulk every night, use batch inference | Day 4: training and inference, inference methods |
| Scenario | Answer | Where it was learned |
|---|---|---|
| A comment-blocking feature where an administrator typed in a list of banned words | Rule-based. Not machine learning | Day 1 |
| A diagnostic aid trained on tens of thousands of X-ray images and their reading results | Deep learning, supervised learning (classification) | Day 1, 2 |
| Enter a sentence and it writes a draft of ad copy | Generative AI | Day 1 |
| Predicting parcel volume as a number by region | Supervised learning (regression) | Day 2 |
| Finding unusual access in access logs with no labels | Unsupervised learning (anomaly detection) | Day 2 |
| A game character learns on its own the moves that raise its score | Reinforcement learning | Day 2 |
| Calculating 10% VAT | Traditional programming | Day 3 |
| A screening score whose basis must be explained to a regulator | Explainable traditional ML (or rules) | Day 3 |
| Preparing a recommendation list for all members every dawn in advance | Batch inference | Day 4 |
| Training 60% / test 58% | Underfitting (high bias) | Day 4 |
| Pair | Words pointing to the first | Words pointing to the second |
|---|---|---|
| Rule-based vs. machine learning | "define conditions," "if-then" | "learn from past data," "patterns" |
| Machine learning vs. deep learning | "tabular data," "features chosen by people" | "many layers of neural networks," "raw images and audio" |
| Classification vs. regression | "whether," "kind," "grade" | "how much," "how many," "how many degrees" |
| Classification vs. clustering | "labeled," "with past outcomes" | "without labels," "similar ones together" |
| Training vs. inference | "build the model," "takes a long time" | "new input," "prediction," "response time" |
| Algorithm vs. model | "learning method," "procedure" | "learned product," "deploy" |
| Validation vs. test | "adjust during training," "choose settings" | "final evaluation," "only once" |
| Overfitting vs. underfitting | "only training is high," "gap," "memorized" | "both low," "too simple" |
| Real-time vs. batch inference | "immediate," "low latency" | "bulk," "periodic," "overnight" |
Next week, building on these fundamentals, we will cover the ML lifecycle (collect → prepare → train → evaluate → deploy → monitor), data, evaluation metrics, and the role of people.
Click a choice to reveal the answer and explanation.
Question 1
Which statement about the relationship between AI, machine learning, and deep learning and about learning types is correct?
Question 2
"Automatically grouping customer data into similar tendencies without answers" and "predicting tomorrow's temperature as a number" are, respectively, which tasks?
Question 3
In a certain business task, the rules are very simple and clear ("amount × fixed tax rate"), and the results must always be 100% accurate. Which judgment is most appropriate?
Question 4
A trained model is nearly perfect on the training data but its performance drops sharply on new data. You are also asked for the definitions of feature and label. Which statement is correct?
Question 5
Putting together what was learned this week, which is the most appropriate judgment for using machine learning well?