Computer VisionBeginner1:22 video6 chapters

Transfer Learning for Vision — lecture notes

Reuse a network trained on millions of images: freeze its layers, add a new head, and fine-tune with only a small dataset of your own.

▶ Watch the animated lecture

0:001. Introduction

Introduction — Transfer Learning for Vision

Training a vision model from scratch needs millions of labelled images. Transfer learning lets you reuse a network that has already learned to see.

0:112. Frozen layers and a new head

Frozen layers and a new head — Transfer Learning for Vision

Take a network pretrained on ImageNet, over a million images. Its layers already detect edges, textures, parts and objects. Freeze them, remove the original classifier, and add a new head for your own classes. Only the head is trained.

0:273. Fine-tuning

Fine-tuning — Transfer Learning for Vision

With a bit more data, unfreeze the last block as well and fine-tune it with a small learning rate. Early layers stay frozen, because edges and textures are useful for almost any task.

0:424. Choosing

With very little data, freeze everything and train only the head. With more data, or data very different from the original, fine-tune some of the top layers too, gently.

0:555. Tips

Match the preprocessing the network was trained with. Use a small learning rate for pretrained layers. Augment your data. And consider modern foundation models, which make very strong starting points.

1:086. Recap

To recap. Start from a pretrained network. Freeze early layers and add a new head. Fine-tune the top layers when data allows. And get strong results even with small datasets.

Key takeaways

Transfer learning reuses features learned on large datasets like ImageNet.
Feature extraction freezes the backbone and trains only a new head.
Fine-tuning unfreezes top layers and uses a small learning rate.
It enables strong models from small datasets.

Check yourself

In feature extraction, which part is trained?
Show answer
Only the new head — The pretrained backbone stays frozen.
Why keep early layers frozen when fine-tuning?
Show answer
Edges and textures are useful for almost any task — Low-level features transfer well across tasks.
What learning rate is typical for fine-tuning pretrained layers?
Show answer
A small one — Small steps avoid destroying useful pretrained features.

Go deeper

Transfer Learning and Fine-Tuning · The AI Lecture Hall
Self-Supervised Vision: SimCLR, MoCo, DINO and Masked Autoencoders · The AI Lecture Hall
CLIP: Connecting Images and Language · The AI Lecture Hall