On the Importance of Being AlphaGo

Victory for #teammachine! Plus a very informative consolation prize for #teamhuman.

This is a response to a couple articles (Beautiful AI and Go, Go and Weakness of Decision Trees) from my friend over at "SE,AG" about the recent match between DeepMind's AlphaGo and World Champion Go player Lee Sedol.:
"Is the Horizon effect something you can just throw more machine learning at? Isn't this what humans do?"
That's what I was going to say in response, but you beat me to it.

Obviously, you could build deeper trees. But the reason these games, even chess, are hard to handle is you can't build larger trees indefinitely. And would it be really satisfying if AlphaGo built ridiculously deep trees, peering into the future more than any human could, in order to defeat a human's uncanny ability to built compact trees using heuristics?

That was why the DeepBlue victory felt a little hollow, as far as I remember, it was mostly tree search with lots of tricks. But AlphaGo uses treesearch, a database of amateur games it has watched and Reinforcement Learning. This means it can really learn from each game it plays, even from playing itself (enter player 0). This is the most exciting thing for me about this victory, that it's a powerful use of RL with deep learning to solve something no one thought would be possible for another decade or two using compute power alone.

Also, on a more pedantic point, AlphaGo doesn't use Decision Trees, which are good for learning how to classify or predict data. Rather it uses Monte Carlo Tree Search, which is an area of research in decision making that simulates forward different scenarios based on the current policy to evaluate how good it is. In many algorithms these trees can be throw away and the policy is updated to take the actual action forward. I think this is what goes on people's head when they play a game like chess or Go. But humans are very good and intuitively picking just the branches with the most promise.

I haven't looked into it in detail yet but I image that part of the Deep Learning element that Alpha Go uses is for building this kind of intuition, so that it doesn't treat all possible choices equally.

How Bad is it Going to Get?

Here's another great article about impending job losses over the next decade due to advanced automation, and it's not just factories anymore. As the tasks that robots and all kinds of  intelligent software systems become more subtle we may approach a point where there there will be a net decrease in total jobs.

"They also said that robots and so-called digital agents will displace more jobs than they create by 2025."

Yet articles like this always include caveats :

"White House economists said they don't have enough information to judge whether increased automation will help or hurt the U.S. economy. For example, new jobs could emerge to develop and maintain robots or other new forms of technology."

This has always been the case in previous ages such as the industrial revolution and the mass production factories of the 1950s. In those eras, the next generation could always learn a new trade and the economy expanded as a whole. I'm an optimist, I believe that the when the children of today grow up they will find ways to add value and 'work' that may be impossible for us to forsee.

But the current revolution requires something those previous ones did not.

"according to the White House, the key is to maintain a "robust training and education agenda to ensure that displaced workers are able to quickly and smoothly move into new jobs."

They're right, current workers in many careers will no longer be able to do the jobs they have. They can retrain, but how fast? These changes are now happening faster than a generation. The user drivers of today putting taxi drivers and dispatchers out of work through technology are funding research into self driving taxis that will lead to the destruction of that entire type of job.

The amount of training required to get the safe jobs that the article talks about takes years and very particular aptitudes. We need to begin to address as a society about all the ways citizens can contribute. Is worn for a wage the only system available? Some Nations will have a better chance of having this discussion than others. But for the nation most at risk to this trend, the United States, I fear that discussion will be incredible hard to even start.

Computers Cannot Make Ethical Choices For Us

I came across this interesting post on the legal and insurance implications of Tesla's upcoming self-driving feature. The new feature would only work for interstate highways in the USA but it would be a big step towards automated driving in cars being widely used in regular driving situations. From a technological perspective it is exciting.

The article notes that fully automatous cars are not going to be common in the near future which they followed with this stunning sentence:

If a circumstance arises where an accident is unavoidable -- say, for instance, a child runs out into the street -- the computers that control the car do not yet have the ethical reasoning to deduce whether they should sacrifice the driver by suddenly swerving away, or run down the child. 

Now, this is an obvious oversimplification of a tricky and possibly unsolvable problem. They linked to some Wired articles which actually provide the horrible ethical dilemma in question and some interesting discussion about what the ethics of a robotic car should be and who should decide them. Partick Lin argues that drivers should not have control over ethical settings on cars to be used to make choices in these situations. The idea of these settings is mostly to save companies from liability and somehow make the driver responsible because of a choice they made on a software screen months or years earlier. Meanwhile, Jason Millar argues the driver should have some input and the situation should be treated as we do informed consent about impossible choices in healthcare. They both make some good points but either way, asking for these preferences up front, rather than in the context of the moment makes it a different moral question. You can't really know the answer until you're there.

Should your car always save you first or pedestrians first, if it computes that there's no way to save both of you? How do you answer such a question? Perhaps more importantly what if the car is wrong? What if the onboard AI has the wrong assessment of the situation and there was a way to save everyone? Then the ethics settings wouldn't really matter would they?

Maybe the question designers need to ask themselves is:
"Do you believe in a no win scenario?"

We live in an age where the machines we interact with daily are becoming intelligent enough to do complex tasks for us, tasks that require training and licensing like driving. We need to be careful not to fool ourselves into thinking these machines are just like us, that they can make human, ethical choices. We've accepted that the fact that machines can play chess better than us doesn't mean they're smarter or that we aren't logical. Driving a car is no different, being able to do it doesn't mean the car can make the hard choices for us.

A machine can only optimize its decisions based on the values it has been given and based on the information it has. In systems like robotic cars, much of that information is uncertain. It has an estimate of what is going on in the world when it is driving and it detects a person jump into the road. It can estimate the odds of getting around the person, the probability of the driver surviving. It can do this faster than us but it does this using its own model and information, not ours.

At this time, no machine can make ethical choices, they can only give a human the information needed to make an ethical choice. Meanwhile, our onboard intelligent computer, our brain, has infinitely more experience and wiring to help us make ethical choices. We also are able to discuss what morality means, justify values based on our lives and those of people we love or based on our beliefs. A machine can do none of this until it has everything we have, until it is truly alive. We humans still get it wrong, quite a lot actually, but right now only humans can make ethical choices. Machines can only help us do it.

So what does this mean for robotic cars? We need to consider another option these two articles didn't, that cars cannot be fully automated until this ethical question is solved. It seems to me this ethical question will only be solved in two ways. We can all agree to accept robotic cars making the wrong choice with a certain frequency, as we accept the infrequent failure of a bridge or plane. Or perhaps someday we will have truly living, intelligent machines which understand morality and values just as well as we do. Although in that situation it might be surprising that such entities would be happy to simply drive our cars for us. Perhaps a general purpose self aware butler AI would do it as one of it's many duties.

Until this happens it would mean the driver of an autonomous robotic car could not take a nap or read a book while the car is driving. They could relax, listen to music or books, or talk on the phone, maybe they could read email on a heads up dash display. But they would need to be alert and aware enough of the situation around them to make a quick ethical choice if the car asks.

This isn't a solution either of course, since the car will need a default to follow if the human fails to respond in time. Or what if the human makes a decision that is clearly the wrong one, will we really blame only them and not the manufacturer of the car for allowing it? I don't know.

The important thing in this discussion is to clearly keep in mind the distinction between machines which make optimizing decisions and humans which can make ethical decisions. These are completely separate entities, and no amount of ethical preference settings will make a car, or any machine, ethical until it knows what morality means. So let's no fool ourselves into thinking it can.