Notes on adversarial network

These days, I heard a lot about un-/semisupervised learning. Ivan Titov’s workshop at the beginning of this month is all about it and now Facebook published a blog post about unsupervised learning for images and videos. They use adversarial networks, i.e. one neural network generates images and another tries to differentiate between authentic and generated images. The first network is trained to fool the other and the second network is trained not to be fooled.

Adversarial network is an interesting idea and I’m amazed by what it achieved. Generating hyper-realistic images is a new power that we couldn’t even imagine some years ago. It will enable great applications (e.g. Photoshop++, virtual reality) and great troubles (e.g. fake evidences, more and more blurry boundary between real and virtual world).

But w.r.t. artificial intelligence, I have my doubts.

Generating realistic images is an artificial task. Try to imagine your house in your mind. Most likely, the image won’t have much details. Now try to draw it on a paper. Try to draw even the simplest object, a coin, a book cover or a pillow. You can always tell apart your drawing and reality with 100% success rate.

But that’s not all, some people can’t imagine at all. No sleep-counting, no sand-and-wave-and-coconut-tree beaches ever cross their mind. Ever. The condition has only been described in recent years and scientists gave it the name aphantasia. I came across this hilarious post from an engineer who was both terrified and relieved when he found out he is aphantasic. Guess what. He is a Director of Product at Facebook.

In short, humans suck at image generation and some can’t do it at all while still being highly intelligent. Trying to master such skills might be a waste of time or at best you will create a totally different kind of intelligence.

Instead, the essence of intelligence is abstract thinking — thoughts that capture things as big as galaxies, as invisible as atoms, as old as the beginning of the universe, as complex as human mind. Generating “realistic” images for such things will be utterly meaningless and useless. As artificial intelligence advances towards more and more abstract levels, adversarial network is likely to be left behind.


Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s