Based on the AIF-C01 exam guideReport an error
Machine learning terms mix English and Korean, and many words look alike. Matching up the pairs with the table below first makes the later explanations easier.
| Term | One-line definition |
|---|---|
| Model | A "bundle of rules" learned from data |
| Training | The process of building a model from data |
| Inference | The process of predicting answers for new data with a trained model |
| Feature | The input information (variable) used for prediction |
| Label | The correct answer the model must predict |
| Overfitting | Memorizing only the training data and doing poorly on new data |
| Underfitting | Too little learning, so it fails to capture even the patterns |
The work of machine learning is broadly split into two stages.
Training is usually done once (or occasionally) as a heavy job, while inference is repeated countless times while the service runs. So the cost and speed requirements for training and inference are different.
| Term | What it is | Cooking analogy | Example |
|---|---|---|---|
| Algorithm | The learning method that finds rules in data | The recipe | "The procedure for finding the line that best fits area and price" |
| Model | The result obtained by running that method on actual data | The finished dish | A formula settled into "price = area × some amount + some amount" |
| Comparison axis | Training | Inference |
|---|---|---|
| What it does | Builds the model | Uses the built model |
| What goes in | Features + labels (large volume) | Features only (a new input) |
| What comes out | A model | A prediction result |
| Frequency | Occasionally, as one heavy job | Continuously while the service runs, each one light |
| What matters | Did it learn well (generalization) | How fast and how cheaply it answers |
How you serve inference depends on "by when the prediction is needed." The exam guide gives four examples: batch, real-time, asynchronous, and serverless. The table below follows the inference options of Amazon SageMaker AI (formerly named Amazon SageMaker).
| Method | When to use | Characteristics | Example |
|---|---|---|---|
| Real-time | When each request needs an immediate answer | An always-on endpoint responds with low latency | Fraud judgment at the moment of payment, product recommendations |
| Batch | Large amounts of data at once, not urgent | Runs a job without a standing endpoint and finishes | Computing churn scores for all customers every night |
| Asynchronous | When inputs are large and processing takes long | Requests are queued and processed, and it can scale down to zero when there are no requests | Analyzing long videos and large documents |
| Serverless | When requests are sparse or uneven | You do not manage instances and pay only for what you use (no charge for idle time) | An internal tool that is called now and then |
| Item | Role | Example |
|---|---|---|
| Floor area, number of rooms, location | Features (input) | 84 m², 3 rooms, Gangnam |
| Actual transaction price | Label (answer) | 900 million KRW |
| Dataset | Purpose |
|---|---|
| Training set | Used to train the model |
| Validation set | Used to check and tune the model during training |
| Test set | Used to evaluate final performance objectively |
💡 Related theory: The reason for splitting data into training, validation, and test sets is to honestly measure whether the model "works well on new data too (generalization)."
These are the representative problem concepts in machine learning.
| State | Performance on training data | Performance on new data | Analogy |
|---|---|---|---|
| Underfitting | Low | Low | Not enough study |
| Appropriate | High | High | Understands and can apply |
| Overfitting | Very high | Low | Only memorized |
A good model works evenly well on both the training data and new data. Balancing between overfitting and underfitting is the core challenge of machine learning, and this is described as "improving generalization performance."
You have to put the two scores side by side to see the state. The training score alone tells you nothing.
| Training data accuracy | Test data accuracy | Diagnosis | How to read it |
|---|---|---|---|
| 62% | 60% | Underfitting | Both are low. It has not learned properly yet |
| 91% | 89% | Appropriate | Both are high and the gap is small |
| 99% | 70% | Overfitting | A 29 percentage point gap. It only memorized and cannot apply |
| State | Common causes | Remedy |
|---|---|---|
| Underfitting | The model is too simple, training is insufficient, there are no useful features | A more expressive model, more sufficient training, better features |
| Overfitting | The data is small or skewed to one side, the model is too complex, it was trained too long | More and more varied data, a simpler model, stop training at a suitable point |
💡 Related theory: The exam guide also asks about this phenomenon with the terms bias and variance. A model that is too simple and consistently misses the correct answer has high bias → underfitting, while one that is swayed sensitively even by small differences in the training data has high variance → overfitting. How well the model has been matched to the data overall is called its fit. Note that "bias" is also used in the sense of fairness, meaning a judgment skewed against a particular group (Week 5), so you must tell from context which one is meant.
The terms learned today connect into a single flow.
[Training data] containing features + labels
↓ (Training)
[Model] completed
↓ (Inference)
New input (features) → predicted label
↓
Evaluate performance with [test data] → check for overfitting/underfitting
Click a choice to reveal the answer and explanation.
Question 1
What is the process of feeding new input data to a completed model to obtain prediction results called?
Question 2
In house price prediction, "floor area, number of rooms, location" are used as input, and "actual transaction price" is the answer to get right. Which terms correctly refer to each?
Question 3
A model achieves very high accuracy on the training data but its performance drops sharply on new test data. Which term best describes this state?
Question 4
What is the main reason for splitting data into training, validation, and test sets in machine learning?
Question 5
What is the state in which learning was insufficient, patterns weren't properly captured even on the training data, and performance is also low on new data?