Unit 3: What did the network actually learn?
About this unit
A low training cost only shows a network fits its data; a held-out test set shows whether it generalises, and looking inside shows its solution rarely matches human ideas.
By the end, you can: Tell generalising from memorising with a test set, and explain what probing a trained network's weights and inputs reveals.
Modules
5 modulesLearn
- Learning or memorising?Article · Test sets, generalization, overfitting, and networks fitting randomly shuffled labels.Not started
- Looking insideArticle · Hidden neurons' weights drawn as images, confident answers on random noise, and interpretability.Not started
- Watch a network overfitExperiment · Train a large network on noisy spiral data and compare training and test loss.Not started
- What a low cost does and doesn't meanReview · Generalization, overfitting, memorisation and what probes of the weights show.Not started