Reversing Neural Networks


What if you could run a deep learning model in reverse? Train up an MNIST classifier, and get the "perfect" image representation of the number 5?


  • You can't inverse the ReLu function, right?

Inversing ReLu

If using tanh instead of relu, the inverse is 12log2((1+x)(1x))\frac{1}{2}*\log_2(\frac{(1+x)}{(1-x)})

I also (occasionally) write an email newsletter! You can subscribe here if you're interested: