Learn AI /How Machines Learn /Transfer learning
Transfer learning
- Intermediate
- How Machines Learn
- Technical readers
Back to the milestone you came from
In five seconds
What it is
Reusing patterns learned on one task or dataset to help a model learn a related task faster and with less new data.
In plain words
Instead of teaching a model from zero, you start with one that has already learned useful patterns elsewhere and adapt it to a related job.
An everyday example
A cook who already knows how dough behaves does not begin from zero when learning a new kind of bread. The ingredients, shape and timing may change, but skills such as judging texture and controlling heat still help. Practice with the new recipe adjusts those existing skills. If the new task were making ice cream instead, much less of the earlier knowledge would transfer.
A radio technician is using explicit human knowledge and can explain each adjustment. A model does not work that way. It stores many numeric parameters shaped by examples, not a neat map it can describe. Also, radio tuning usually adjusts one signal at a time, while transfer learning may involve changing millions of parameters across many training steps.
Remember thisTransfer learning gives a related task a learned starting point. It can reduce the new data and training required, but its value must be tested because source patterns do not always transfer well.
Understand it
Understand it
Why it matters
Transfer learning affects time, cost and risk. If it works well, an organisation can build a useful model with less labelled data, less computing, and shorter development time than training from scratch. That matters when expert labels are expensive, such as in medicine, law, agriculture or multilingual document analysis. It also changes procurement and evaluation decisions: you need to ask what the model was trained on before, whether that earlier source is relevant to your task, and whether reused patterns might carry bias or mismatch. A model that transfers well on one problem may transfer poorly on another, so decision-makers should treat it as a practical strategy, not a guarantee.
Training from scratch versus transfer learning
Show why starting from a pre-trained model can reduce data needs and training effort for a related task.
01
Path A: start from scratch
A new model begins with random parameter values, so it must learn both general patterns and task-specific patterns from the target dataset.
Top row: box labelled 'Randomly initialised model' leading to a wide training box labelled 'Large target-task training' and then to 'Target model'.
02
Path B: reuse prior learning
A model first trained on a broader or earlier task already contains useful general patterns.
Bottom row begins with box 'Pre-trained model from earlier task' using a filled colour distinct from the top row.
03
Adapt to the new task
The pre-trained model is updated on the target dataset. Some layers may stay fixed, while others are changed.
Arrow from 'Pre-trained model' to a smaller training box labelled 'Adapt on target task'; add a side note icon 'some layers frozen, some updated'.
04
Compare effort and fit
For related tasks, the transfer path often needs less labelled data and fewer training updates to reach useful performance.
Place a bracket or annotation under both rows: top says 'more target data and training', bottom says 'often less target data and training for related tasks'.
A side-by-side diagram compares two paths to the same target task. Read left to right. The top path starts with an untrained model and requires large task-specific training. The bottom path starts with a pre-trained model, then adapts it with a smaller task-specific dataset.
Hover or tap each stage to reveal a short note: what is reused, what is newly trained, and when the approach may fail because the tasks are too different.
Describe this without the diagram
This diagram compares two ways to build a model for a target task. In the first path, a randomly initialised model is trained only on the target dataset, so it must learn everything from that data. In the second path, a pre-trained model from an earlier related task is reused and then adapted on the target dataset. The key message is that prior learned patterns can give a useful starting point, so the second path often needs less target data and fewer updates when the tasks are related. A note adds that some parts of the model may be kept fixed while others are updated.
With reduced motion: A static two-row comparison is shown instead of animated arrows. The top row shows training from scratch. The bottom row shows reuse of a pre-trained model followed by adaptation on the target task, with a note that transfer learning helps most when tasks are related.
Transfer learning starts with a model that has already been trained on a large earlier task. That earlier stage usually teaches the model useful patterns, such as edges in images, word relationships in text, or general signal structure in audio.
Then a second task is introduced. Instead of training a new model from the beginning with random starting values, developers reuse some or all of the earlier model and adapt it to the new task. The new input is data for the second task, and the output is a model that is specialised for that task.
Inside the model, the main change is that existing internal parameters, often called weights, are kept as a starting point rather than thrown away. Sometimes only the final layer is replaced and trained. Sometimes many layers are updated. This later adaptation step is often called fine-tuning, but transfer learning is the broader idea of reusing prior learning at all.
This works best when the first and second tasks are related. A model trained on general language can often be adapted for classifying support tickets. A model trained on many everyday images can often be adapted for identifying crop diseases. If the tasks are too different, the reused patterns may help little or may even hurt performance.
How “Adapting an experienced cook's skills” maps onto the real thing
| In the story | In the AI system |
|---|---|
| Earlier map of the FM band | Model trained on an earlier, broader task |
| Starting near likely frequencies | Initialising the new model from learned weights instead of random values |
| Small adjustments to clear the signal | Adapting the model on the new task with additional training |
| Different local station on another radio | A related but distinct target task or dataset |
A worked example
An information management officer at a regional hub is preparing an archive of public situation maps for upload to a new portal. Only a small labelled sample is available for sorting the maps as access, displacement or service coverage products.
Starting from: Start with an image model pre-trained on a broad image collection, then adapt it using the labelled situation maps.
- Replace the model's original output layer with a classification head for the three map types.
- Freeze the lower layers, which capture general visual features, and train the new classification head on the labelled maps.
- Unfreeze selected upper layers and update them with a low learning rate so they can adapt to map layouts, symbols and legends without discarding useful general features.
- Evaluate the adapted model on held-out maps and compare it with a model trained from random initial weights on the same training sample.
Result: The comparison establishes whether the source model's visual representations transfer to this task. Better performance on held-out maps would support using the adapted model, subject to review of errors across different templates. If it does not outperform the baseline, pre-training has not provided the expected benefit.
In a United Nations workplace
During a records retention review, a records officer in a specialised agency technical division must propose retention categories for entries in the records inventory before a note for the file goes to the chief of section. A language model pre-trained on general documents is adapted using approved, previously categorised inventory entries.
The pre-trained model already represents common wording and document structure. Fine-tuning selected weights on the division's examples can adapt those representations to its records vocabulary with less labelled material than training a model from scratch.
Before you try thisRecords containing personal, HR, procurement-sensitive, security-sensitive or otherwise confidential material must not be uploaded to or used to train a tool not approved for that material. Use an approved environment, minimise and de-identify data where possible, apply access and retention controls, and have the records officer verify every proposed category against the applicable retention schedule.
Where you may hear it
- A technical review before an ICT governance board
- A vendor demo during a procurement meeting
- A model evaluation session at a regional hub
- An inter-agency working group discussing a low-data pilot
Avoid confusion
Avoid confusion
Often confused with
Fine-tuning
Updating a pre-trained model's parameters with additional examples to improve performance on a narrower task, style or domain.
Pretraining
A large-scale training stage where a model learns reusable patterns from data, producing a general-purpose starting point that later steps can adapt for specific tasks.
Inference
Using a trained model on new input to produce an output such as a score, category, translation or generated text.
Common misconceptions
Common misconception: Transfer learning and fine-tuning are the same thing.
Correction: They are related but not identical. Transfer learning is the broader strategy of using knowledge from an earlier task for a new one. Fine-tuning is one common method for doing that by continuing training on the new task. Some transfer learning setups reuse a model as a fixed feature extractor with little or no further updating.
Common misconception: A pre-trained model always performs better than training from scratch.
Correction: It often helps, but not always. If the original task is too different from the new task, the reused patterns may be unhelpful or even harmful. Performance depends on task similarity, data quality, model size and how adaptation is done.
Common misconception: Transfer learning means no new labelled data is needed.
Correction: Sometimes very little new data is enough, but usually some target-task data is still needed to adapt and evaluate the model properly. You also still need validation and test data to check whether the transferred model actually works for your setting.
Common misconception: Because the model learned on a huge dataset before, it is automatically suitable for any domain.
Correction: Large earlier training can provide broad patterns, but domain differences still matter. A model trained on everyday photographs may not transfer cleanly to satellite imagery. A language model trained on public web text may not match formal legal or humanitarian reporting language.
Quick check
Which case best matches transfer learning?
Transfer learning uses prior learning from an earlier task or dataset to help with a new related task. Option two does that. Option one is training from scratch, option three is inference, and option four describes clustering.
Go deeper
Go deeper
The simple version
Instead of teaching a model from zero, you start with one that has already learned useful patterns elsewhere and adapt it to a related job.
The technical version
Transfer learning reuses parameters, representations or features learned from a source task to improve learning on a target task. In practice, a model is often pre-trained on a large dataset, then adapted using target-domain data by freezing some layers, replacing task-specific heads, and updating selected weights through gradient-based optimisation. The expected benefit comes from shared structure between source and target distributions, so the model does not need to relearn low-level or general-purpose patterns from random initialisation. The term is broader than fine-tuning: it can include fixed feature extraction, multi-stage training and other ways of carrying information from one learning problem to another. Its success depends on relatedness between tasks, representation quality, label availability, and evaluation on separate validation and test sets.
Reviewed external resources
EncyclopaediaTransfer learning (opens in a new tab)Wikipedia
This Wikipedia page gives a broad overview of transfer learning concepts and terminology, and it suits readers who want a general reference before going deeper.
Intermediatelink checked 2026-07-31
CourseMachine Learning Crash Course (opens in a new tab)Google
This course material gives practical machine learning context that helps readers place transfer learning within real workflows, and it suits learners who prefer structured lessons.
Intermediatelink checked 2026-07-31
VideoStatQuest (opens in a new tab)StatQuest with Josh Starmer
This StatQuest video is likely to give an intuitive step-by-step explanation in plain language, and it suits readers who learn well from visual teaching.
Starterlink checked 2026-07-31
VideoVideo explanations: Transfer learning (opens in a new tab)YouTube search
This YouTube search results page helps readers compare multiple video explanations of transfer learning, and it suits people looking for a teaching style that fits them.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Transfer learning (opens in a new tab)Reddit search
This Reddit search results page helps readers see common questions, confusions and practitioner discussions about transfer learning, and it suits readers who want informal community perspectives.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
Transfer learning gives a related task a learned starting point. It can reduce the new data and training required, but its value must be tested because source patterns do not always transfer well.