It is very neat. There is no obvious way to apply these methods to nets with many hidden units (which means most useful deep learning nets). The visualization techniques are excellent though. It’s a great blog post, and we’ll see if it leads to some great research.
Since one big problem with neural nets is their lack of analyzability, this geometric approach to deep learning neural networks seems probably useful.
It is very neat. There is no obvious way to apply these methods to nets with many hidden units (which means most useful deep learning nets). The visualization techniques are excellent though. It’s a great blog post, and we’ll see if it leads to some great research.