Transfer Learning and Pretraining
For engineers who can train a basic neural network and now need good results from a few hundred or a few thousand labelled examples. You will be able to adapt pretrained image and text models with feature extraction and staged fine-tuning, cope with small datasets and domain shift, and measure the results honestly.
Course Content
Why pretrained weights beat training from scratch on small data, how to decide what to freeze and what to fine-tune, and what to do when deployment data drifts away from training data. 3 lessons, about 55 minutes.
Hands-on transfer learning with real libraries: choosing and loading image backbones, fine-tuning BERT-family text models, and getting trustworthy results from very small datasets. 3 lessons, about 50 minutes.
A capstone build: fine-tune a pretrained ResNet-50 on Oxford Flowers-102 in measured stages, then analyse its errors instead of stopping at one accuracy number. 1 lesson, about 20 minutes.