A company wants to apply the knowledge of a model trained on one task to a related but different task, to build a new model efficiently with little data. The aim is to reduce the cost of data collection. Which technique represents this idea?

1 / 1
Select an answer
CorrectB

Explanation

Question Overview

Select the technique that applies existing knowledge to another task.

Requirements to satisfy
  • 1「the knowledge of a model trained on one task」Reuse the learned knowledge
  • 2「to a related but different task」Reuse for another task = transfer learning
Per-option explanation
AIncorrect

Pre-training

Pre-training is the stage that acquires general knowledge from scratch with a large amount of data.

It differs from the idea of reusing existing knowledge for another task to build with little data, so it is incorrect.

BCorrect

Transfer learning

This is correct. Transfer learning is the idea of applying knowledge learned on one task to another related task to build a model efficiently with little data. Fine-tuning a foundation model is one form of this.

CIncorrect

Regularization

Regularization is a technique that suppresses overfitting during training.

It is not the idea of reusing knowledge from an existing task for another task, so it is incorrect.

DIncorrect

Data augmentation

Data augmentation is a technique that inflates training data by processing existing data.

It is not the idea of reusing knowledge from an existing task for another task, so it is incorrect.

Key Takeaway

Remember the idea of the correct answer, 'transfer learning.'
- Applies knowledge learned on one task to another related task to build a model efficiently with little data.
- Fine-tuning a foundation model is one form of transfer learning.
It is a different concept from pre-training (general-purpose learning from scratch), regularization (suppressing overfitting), and data augmentation (inflation).

Related Links