[waffl]

While I think this is beautiful, conceptually, I really am a bit terrified of the potential of this in reverse (the neural network for processing/understanding an image). With Google releasing their 'Photos' app, this network is about to get a direct pipeline for machine learning imagery to accelerate everything – my main fear would be the potential for this technology to be employed by weaponized drones able to scan a scene (with, eventually, incredibly high resolution cameras and microphones that far surpass human capability) and identify every single object/person in realtime (also at a rate that humans are incapable of).

Of course, there is great utility to be had as well, it just scares me to think about what could be done with this technology, in a mature form, if used for violent purposes.

[visarga]

This will happen for sure. Such super-perceptive computers will oversee our every movement.

Computers can already understand our emotions in writing, voice and from the expression on our faces, they can also estimate pose and understand your movements. They can label thousands of kinds of objects. And they're just starting.

They can also build neural nets 10x smaller by compressing a larger neural net while maintaining most of accuracy. That means once a problem such as vision or speech has been solved with a huge net, it can be transferred in a smaller, more efficient net.

[aoeuasdf1]

> They can also build neural nets 10x smaller by compressing a larger neural net while maintaining most of accuracy. That means once a problem such as vision or speech has been solved with a huge net, it can be transferred in a smaller, more efficient net.

This is known as "dark knowledge". Slides from Geoff Hinton: http://www.ttic.edu/dl/dark14.pdf