Transfer Learning
Starting from a model that already learned something general, instead of from random numbers — why small teams can build real AI.
When not to use it
- When your domain is genuinely unlike the source. Medical scans, satellite imagery, industrial sensor data — transfer from internet photos helps far less than the reputation suggests, and can hurt.
- When you have plenty of data. Above a certain scale you can train something better suited than anything you'd inherit.
- When the pretrained model's licence or training data is a problem. That's a legal question you inherit along with the weights, and it doesn't announce itself.
Reach for something else instead
- Training from scratch when your domain is alien and your dataset is large.
- Feature extraction only — freeze everything, use the outputs as inputs to a simple classifier. Cheap, fast, hard to overfit.
- In-context learning with a foundation model, which adapts behaviour with zero training.
Sources & further reading
- Yosinski et al. (2014), How transferable are features in deep neural networks? — measures which layers transfer and which don't. The paper that made this concrete.
- Howard & Ruder (2018), Universal Language Model Fine-tuning for Text Classification — discriminative learning rates and gradual unfreezing.
- Raghu et al. (2019), Transfusion: Understanding Transfer Learning for Medical Imaging — where transfer helps less than assumed, and why.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Fine-tuning everything at full learning rate on a small dataset, destroying the features you came for.
- Preprocessing differently from the original training. The model expects that exact normalisation and simply performs worse without it.
- Assuming transfer always helps. Negative transfer is documented, and the only way to know is to compare against a scratch baseline.
At a glance
FieldDeep Learning
Core ideareuse general features, retrain the specific ones
Freezeearly layers
Adaptlater layers, low rate
Risknegative transfer, catastrophic forgetting
DifficultyIntermediate
Flashcards for this concept
Question
—
Answer
—
1 / 4
Often compared with
Transfer learning vs. training from scratch — inheriting general features vs. buying them with data and compute.