A tiny neural network that shows the strenghts and weaknesses of deep learning.
Trained on the CIFAR-10 dataset, this neural network is used to show what a deep learning can and can not do.
- It classifies off-center images correctly.
- It classifies a lego model of a firetruck correctly.
- It classifies a black hole as a ship correctly. Proving once and for all that black holes are ships between dimensions.
- It thinks a fictional frog is a deer, probably because it detects an antler silhouette.
Deep learning models (or any machine learning model for that matter) can only do what they were trained for. Put them even a micrometer out of context, and the weirdest stuff will appear.