Course Wiki

Transfer Learning

Reusing a pretrained convolutional network's learned features instead of training an image model from scratch.

Transfer learning is the module’s answer to small datasets: take a large convolutional network pretrained on ImageNet, freeze its convolutional base, and train only a small classification head on your own images. The pretrained layers already encode useful visual features — edges, textures, shapes — so the head needs far fewer examples to converge.

The course applies this with Xception and similar architectures in Keras and PyTorch, and it is the standard technique in learner capstones: the published waste-classifier case study fine-tunes an Xception model to 93% test accuracy on roughly 15,000 images.

Taught in

• Machine Learning Zoomcamp

  • Module 8: Neural Networks and Deep Learning
    • Transfer learning
    • Explore more ## Related concepts

• Neural Networks

• Serverless Deployment

• Gradient Boosting

• Docker