← AVIATION GLOSSARY
A B C D E F G H I J K L M N O P Q R S T U V W X Y Z
Aviation glossary

What is Transfer Learning?

Machine Learning · Computer Vision 8 languages
Definition
Transfer learning is a machine learning method that uses a model trained on one task to improve learning on a different task.

Transfer learning is a machine learning method that uses a model trained on one task to improve learning on a different task. Pan and Yang define it with a source domain and task and a target domain and task. The domains or the tasks differ.

How it works

A model is first trained on a large source dataset. For image models this is often ImageNet, which has about 1.2 million training images in 1000 classes. The model is then reused for the new task, which has less labeled data. There are two common ways to reuse it:

  • Feature extraction. Remove the last classification layer, keep the rest frozen, and train a new linear classifier on the output.
  • Fine-tuning. Replace the classifier and continue training some or all of the pretrained layers. The pretrained weights get a smaller learning rate than the new classifier.

Few teams train a whole network from scratch, because large datasets are rare. For a small target set that looks like the source data, the Stanford CS231n notes advise feature extraction. A large target set allows fine-tuning of the whole network.

Why the features transfer

Yosinski and co-authors tested this layer by layer. Features in the first layer, such as color blobs, are general. Later layers are more specific to the source task. The more the target task differs from the source task, the less the features transfer. Starting from transferred weights can still help after fine-tuning.

Self-supervised backbones

A backbone does not need labels to be useful. DINOv3 from Meta AI is trained without manual labels on a curated set of 1,689 million images. Its largest model has 7 billion parameters. The authors report that it gives strong results on many tasks without fine-tuning, and in most of their tests the backbone stays frozen. See also the DINOv3 page in this glossary.

Airfield inspection

Airport pavement condition index (PCI) surveys follow ASTM D5340, as FAA AC 150/5380-7B states. The circular does not mention machine learning. A model built with transfer learning can only find defects in images. It does not change the survey method.

Frequently Asked Questions

What is transfer learning?
Transfer learning is a method that uses the knowledge from a source task to improve learning of a target task. The source and target can differ in their data, in their task, or in both. In computer vision, the usual source is a network trained on a large image set such as ImageNet.
What is the difference between feature extraction and fine-tuning?
In feature extraction, the pretrained network stays frozen and only a new linear classifier is trained on its output. In fine-tuning, the classifier is replaced and training continues through some or all of the pretrained layers. Fine-tuning uses a smaller learning rate for the pretrained weights than for the new classifier.
Why not train a network from scratch?
Training a whole network from scratch needs a large dataset, and few projects have one. A pretrained network already holds general features, so a smaller dataset can be enough for the new task.

References

pan-yang-2010
A Survey on Transfer Learning S. J. Pan and Q. Yang, IEEE Transactions on Knowledge and Data Engineering, vol. 22, no. 10, 2010
cs231n-transfer
Transfer Learning, CS231n Convolutional Neural Networks for Visual Recognition Stanford University
yosinski-2014
How transferable are features in deep neural networks? J. Yosinski, J. Clune, Y. Bengio, H. Lipson, arXiv 1411.1792
ilsvrc-2015
ImageNet Large Scale Visual Recognition Challenge O. Russakovsky et al., arXiv 1409.0575 (International Journal of Computer Vision, 2015)
dinov3-2025
DINOv3 O. Siméoni et al., Meta AI, arXiv 2508.10104
faa-ac-150-5380-7b
AC 150/5380-7B, Airport Pavement Management Program FAA

This is a reference entry. TarmacView provides drone inspection services and doesn't sell the equipment or materials described in this glossary.

Get Started Need your airfield lighting verified? TarmacView inspects PAPI, runway and approach lighting by drone, between aircraft movements, and delivers a compliance report built for your regulator. Pavement and obstacle inspection are in development.
Request an inspection quote See how a drone inspection works
Keep reading

Learn more

03 PAGES
DINOv3 Vision Transformer
glossary

DINOv3 Vision Transformer

DINOv3 (self-DIstillation with NO labels v3) is a self-supervised vision transformer (ViT-B/16) pretrained on 1.7 billion images, producing high-quality 768-dimensional embeddings that capture fine-grained texture and structure. TarmacView models in development use DINOv2 and DINOv3 backbones, for different purposes. Covers DINO training, ViT architecture, fine-tuning for domain tasks, and…

Domain Adaptation
glossary

Domain Adaptation

Domain adaptation adapts machine learning models trained on a source domain, such as specific pavement types, lighting conditions, or datasets, to perform reliably on different target domains without full retraining. Critical for infrastructure inspection when deploying from road pavements to airport runways or from sunny to overcast conditions.

Data Augmentation
glossary

Data Augmentation

Data augmentation synthetically expands training datasets by applying image transformations, rotation, flipping, color jitter, blur, noise, cropping, to improve model robustness to variations in lighting, orientation, and image quality. For infrastructure inspection, domain-specific augmentations (perspective transforms, shadow simulation, weather effects) are critical. Covers augmentation…