When studying machine learning, the same concepts get referred to with a mix of English and Korean terms, and many words look similar, which can be confusing. Today we will precisely sort out the foundational vocabulary you must know for the AIF-C01 exam and for real-world conversations.
The terms we will cover today are: model, training, inference, feature, label, datasets (training/validation/test), and overfitting and underfitting.
| Term | English | One-line definition |
|---|---|---|
| Model | Model | A "bundle of rules" learned from data |
| Training | Training | The process of building a model from data |
| Inference | Inference | The process of predicting answers for new data using a trained model |
| Feature | Feature | Input information (variables) used for prediction |
| Label | Label | The answer the model must get right |
| Overfitting | Overfitting | A state of memorizing only the training data and performing poorly on new data |
| Underfitting | Underfitting | A state of insufficient learning where even the patterns aren't captured |
A model is the artifact learned from data. Think of it as a "learned bundle of rules" that takes input and produces output. For example, a function that takes a house's floor area as input and outputs an estimated price is a model.
A model itself starts out as a blank state that knows nothing, and it gradually becomes smarter through training.
Machine learning work is broadly divided into two stages.
By analogy, training is "the process of studying," and inference is "solving exam problems with what you've learned."
💡 Related theory: Training is usually performed heavily once (or occasionally), while inference is repeated countless times during service. That is why the cost and speed requirements of training and inference differ from each other.
Using the house price prediction example:
| Item | Role | Example |
|---|---|---|
| Floor area, number of rooms, location | Features (input) | 84㎡, 3 rooms, Gangnam |
| Actual transaction price | Label (answer) | 900 million KRW |
Training requires both features and labels, while inference feeds in only features to predict the label (answer).
To build a model properly, the available data is usually split into three parts.
| Dataset | English | Purpose |
|---|---|---|
| Training data | Training set | Used to train the model |
| Validation data | Validation set | Used to check and tune the model during training |
| Test data | Test set | Used to objectively evaluate final performance |
The key point is that test data is never used for training. Evaluating with data used for training is like "taking an exam with questions you've already seen," so performance gets inflated.
💡 Related theory: The reason for splitting data into training/validation/test is to honestly measure whether the model "also works well on new data (generalization)."
These are the most frequently encountered problem concepts in machine learning.
A state where the model has memorized the training data too well, and as a result performs poorly on new data. It is likened to "a student who memorizes past exam papers wholesale and can't solve variations."
A state where the model has learned too little and failed to even capture the patterns. It is likened to "a student who studied so little that they get even the basic questions wrong."
| State | Training data performance | New data performance | Analogy |
|---|---|---|---|
| Underfitting | Low | Low | Not enough studying |
| Just right | High | High | Understands and can apply |
| Overfitting | Very high | Low | Only memorized |
💡 Related theory: A good model performs consistently well on both training data and new data. Striking the balance between overfitting and underfitting is the central challenge of machine learning, and this is described as "improving generalization performance."
The terms learned today don't stand alone — they 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
If you can picture this flow in your head, the more complex content in later weeks will be much easier to understand.
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?