Why deep neural networks don’t actually think

medium.com

5 points·sshamoon··2 comments

This article (like most articles on this topic) doesn't define what "think" means. There's no shame in that: nobody has a good definition. But without a good definition of X, you can't make reliable conclusions about what is or isn't X.

The logic seems to be: Thinking is mysterious [true]. Neural nets aren't mysterious [true]. Therefore neural nets must not be thinking [?].

That's a fallacy. Compare: The location of the buried treasure is mysterious. The place I'm about to dig isn't mysterious. Therefore, the treasure must not be where I'm about to dig.

It's a fallacy because as soon as you find the treasure, it's location isn't mysterious any more. Mysteriousness isn't an inherent property of things, it's a statement about our own limited knowledge which changes over time.

The same will be true of thinking (for any definition of that term you might care to use). When we figure out how to make artificial systems that think, thinking will no longer be (completely) mysterious.

There is (probably) a definition of "think" where neural networks actually thing. There is also (probably) a definition of "think" where neural networks don't actually think. There are articles advocating both positions, and neither side explicitly states their definition of "think". So they're really arguing about what the right definition is, but in disguise, since neither side states their definition, nor gives reasons why that definition is better than the other side's. As a result, the "discussion" is pretty much useless.

Reply on news.ycombinator.com