Adversarial Training
Learn about adversarial training of images, what it is, and how it is done.
In 2014, Ian Goodfellow presented a different kind of network architecture. It wasn’t a bigger, wider, or deeper version of other neural networks. It didn’t use fancier activation functions or more advanced optimization techniques. It was structurally different.
Let’s approach this idea one small step at a time.
The simple image classifier: cat vs. not cat
The following picture shows a neural network that learns to classify images as cats or not cats.
-
If the image fed to the network is a cat, the output value should be 1, for true.
-
If the image isn’t a cat, the output should be 0, for false.
This architecture is not too different from our MNIST example. The only difference is that this classifier outputs a single value, not ten.