Faking Feature Space Representations

 

So i was reading this multi-modal neuron feature representation paper (neural net stuff), and they were generating some wild output images. So i wanted to know how they were doing their back-prop optimization from a feature vector in their neural net to the actual images they were showing.

And then i looked at the PyTorch code for that part. And i realized you can get there way, way simpler than what they are doing.  Spent 5 minutes in Studio Artist V5.5 dialing in something close. Their facets are better, but i think you could get very close just using existing Studio Artist features.

So this would be visualizing what a clown neuron represents. Here's another slightly different approach.