THE CORE AIM OF VOCABULARY MASTERY · TRANSFER LEARNING VOCABULARY · PRETRAIN → TRANSFER → ADAPT → FINE-TUNE → EVALUATE
Transfer learning vocabulary is the language used to describe reusing knowledge learned by one machine-learning model for a new task, dataset or domain. Terms such as pretrained model, source task, target task, feature extractor, fine-tuning, frozen layer, domain shift and negative transfer matter because modern AI often begins with an existing model rather than learning everything from scratch.
The core aim of vocabulary mastery for transfer learning vocabulary is adaptation clarity. Learners should be able to explain what knowledge the original model learned, what new task it is being adapted to, which parameters are reused or updated and how evaluation shows whether transfer actually helped.
This page is the Transfer Learning Vocabulary owner inside the eduKateSG Vocabulary hub. For model foundations, use Machine Learning Vocabulary. For neural-network training, use Deep Learning Vocabulary.
Central proposition: Transfer-learning vocabulary is mastered when the learner can state what is reused, what is changed, why transfer should help and how that claim is tested.
The 60-Second Transfer Learning Vocabulary Router
- Origin: pretrained model, source dataset, source task.
- Destination: target dataset, target task, target domain.
- Reuse: representation, feature extractor, checkpoint.
- Adapt: freeze, unfreeze, fine-tune, learning rate.
- Shift: domain shift, distribution shift, mismatch.
- Evaluate: baseline, generalisation, negative transfer.
Pretraining and Fine-Tuning Are Different
Pretraining learns broadly useful representations or capabilities from an initial dataset or objective. Fine-tuning continues training a pretrained model for a narrower task or domain. The second stage depends on knowledge acquired in the first.
A Worked Example: Frozen Layers
A frozen layer keeps its learned parameters unchanged during adaptation. Freezing can preserve useful representations and reduce training cost, while unfreezing more layers allows deeper adaptation at the risk of overfitting or forgetting useful prior knowledge.
A Worked Example: Feature Extractor
A pretrained network can act as a feature extractor: early or intermediate representations are reused while a new task-specific output layer is trained. This is common when the target dataset is relatively small.
Domain Shift and Negative Transfer
Domain shift means the target data differs meaningfully from the source data. Negative transfer occurs when reused knowledge harms performance compared with an appropriate baseline. Transfer is therefore a hypothesis to test, not an automatic improvement.
How to Learn Transfer Learning Vocabulary
- Compare training from scratch with a pretrained model.
- Identify source and target tasks.
- Freeze and unfreeze layers conceptually.
- Track which parameters change.
- Measure performance against a baseline.
- Look for domain mismatch.
- Explain why transfer helped or failed.
Common Transfer Learning Vocabulary Mistakes
Calling every pretrained model transfer learning
Repair: identify the actual target adaptation or reuse.
Confusing fine-tuning with prompting
Repair: fine-tuning changes model parameters; prompting changes input context.
Assuming transfer always improves performance
Repair: evaluate against a relevant baseline and watch for negative transfer.
Frequently Asked Questions
What is transfer learning vocabulary?
It is the language used for pretrained models, source and target tasks, feature reuse, freezing, fine-tuning and domain adaptation.
What is fine-tuning?
It is continued training of a pretrained model for a target task or domain.
What is negative transfer?
It is when transferred knowledge makes target-task performance worse than an appropriate alternative.
The Transfer Learning Vocabulary Standard
Mastery means explaining what prior knowledge is reused, what changes during adaptation and what evidence shows the transfer is useful.
That is the standard: adaptation language precise enough to distinguish reuse from improvement.