Deep learning, deep knowledge, and super-intelligence

Originally published on Tumblr.

To hear the press go on about deep learning you’d think we were nearly at the end of history.

We don’t need humans to see, listen, and understand, because computers can do it better! We don’t need to write software anymore, because we can just teach computers what they need to know. In fact, we probably don’t need to teach them at all. They can teach themselves!

How is all this possible? Deep learning!

Maybe there’s a little hype in the news, but it’s only a question of time before computers really are able to sense and understand the world better and faster than we can.

Let’s back up a bit and try to put it in perspective.

What is deep learning?

It’s a way of programming computers so that they work the way we think our brains work. Incredibly, this technique works very well. Using deep learning, computers can now recognize objects almost as well as humans. (Actually, last week Microsoft announced that their latest system does better than humans!) Scientists are also using deep learning to recognize speech, translate text, and myriad other things. Some of these solutions work very well. Some need to work better.

So why are people this excited by an algorithm?

Because some of the things it’s good at — like image and speech recognition — are things that, until recently, only humans could do well. And having machines do these things makes a lot of new things possible. Without amazing vision, cars can’t drive themselves. Without great hearing, computers are bound to input from keyboards and mice, when speech makes a lot more sense. These are big deals.

But deep learning has limitations. To learn about the world, a deep learning system needs to process a lot of real world data. Without data, the algorithm is useless. Even a trained system can easily get confused in the face of new data. A deep learning speech system trained on American speech, for example, might understand a New Yorker, but have a very hard time understanding a South African. We still haven’t figured out how get a deep learning system to adapt to these variations without retraining.

Until recently we’ve had to teach a deep learning system by telling it about the training data. So, for example, if we wanted it to learn to recognize bridges, we’d show it hundreds of thousands of bridge pictures from different angles, and tell it that they’re bridges. A few months ago, however, Google announced that a deep learning algorithm they developed learned to recognize cats on YouTube without being told what a cat was. That’s a startling advance that could have major implications.

As exciting as deep learning is, it should give us pause. Consider the following.

We are recorded on CCTV every day. Most of us don’t care, because we know that no one can possibly review all those tapes. Except in cases of criminal investigations most of them are archived and never seen by a human. But computers are fast, and getting faster. Before long computers running recognition software will be able to watch every video ever recorded, as well as every photograph ever taken, and find you in all of them. Before long deep learning systems will be able to listen to every phone call ever recorded and not only identify you, but also understand what you said. And all of that information could easily be correlated to financial transactions, as well as educational, work, and medical records. This will create a very accurate picture of your life, a picture more accurate than any available today. This is deep knowledge.

This new kind of artificial knowledge will allow an individual, corporation, or government to learn vastly more about another person than is possible today. In fact, this knowledge will be so detailed and precise that it will far outstrip our own memories in terms of granularity and accuracy. The implications are worrisome. For example, this deep knowledge will allow law enforcement to retrospectively find crimes that were never reported. It will allow an employer, or a friend learn about highly embarrassing things you did at any time in your life, things you wish you could forget.

Without very significant restraints on the access and use of this knowledge, no one will be safe from ridicule, ostracism, or prosecution.

Many people believe that deep learning and deep knowledge will inevitably lead to super-intelligence, a form of intelligence that surpasses human intelligence by an unfathomable degree. As many people are pointing out, super-intelligence changes everything. Once we humans are no longer at the top of the heap, we have no idea what will happen.

I’m happy that people are worrying about the implications of artificial intelligence outstripping our own. But first, it’s important that political minds spend time worrying about the implications of artificial intelligence, which will quickly destroy a large percentage of today’s jobs. It’s also important that legal minds spend time worrying about individuals’ relationship with society once everything those individuals do is “known”.

I don’t know if super-intelligence is inevitable, but deep learning and deep knowledge are. Let’s apply some ordinary intelligence to to the problems we’re certain to face.