What is Intelligence: The True Power of Deep Reinforcement Learning

THE TRUE POWER OF DEEP REINFORCEMENT LEARNING

It is very easy to be swayed by philosophical thoughts when contemplating artificial intelligence. I too have my personal AI fears, and it has much to do with automated weaponry than it has to do with superintelligence.


If one sees how systems powered by DeepRL (Deep Reinforcement Learning) efficiently go after their goals. shivers will indeed be sent down your spine. I am just shocked at how efficiently it goes after its goals as I ponder how such systems could be converted into weaponry.

Is that real intelligence in action? If acting rationally and efficiently towards attaining a goal is taken as a definition of intelligence then yes DeepRL systems show some kind of intelligence.

I am also impressed by the algorithms that beat the best Go players. In some of my writeups I may have come up as bashing the achievements of game playing AI, but in reality, I am not bashing the algorithmic muscles that have gone into making such great achievements.

What I try to do in my writings is to bring sanity to a field that is almost being consumed by hype. I am an AI researcher and will be much affected if the current AI enthusiasm ends in some kind of bubble burst scenario.

What I am concerned about is correcting the interpretations of current progress in AI, putting the focus on where it needs to be and reducing the hype to allow us to realize where we truly are on the path of AI development and allowing us to develop realistic goals that could clearly take us to the next level.

However, the other camps whose brains extrapolate way too much from very little information would like to interpret current progress in game playing AI as a direct reflection that we are about to arrive at superintelligence.

Apart from the fact that DeepRL system can achieve their goals using the most efficient path possible, far more efficient than humans can hope to achieve using reasoning alone, and we can clearly see how easily this could be used to drive all kinds of weapon systems that can effectively eliminate targets more efficiently than humans.

The most disturbing aspect of these systems is the difficulty of properly defining a goal for them. Like the example Lex Fridman usually refers to in his MIT AI lectures, a DeepRL system with improperly defined goals will try its best to achieve whatever interpretation of that goal it can come up with. In the last sentence, I was trying so hard to use the statement: whatever interpretation of that goal it can “think” of. Because this is not really thinking in any form and is just some kind of blind optimization.

In the example Lex uses, there is some agent who is given a goal of winning in a game by obtaining the highest amount of points possible. These points are obtained along the wayside, finishing the game is taken for granted as implicit. A human would take as many points as possible but will still try to reach the end of the game while a DeepRL system with ill-defined goals might try to optimize for points alone and will remain spinning around obtaining points in an area of high point density. Overall it attains a lot of points without ever getting to complete the game.

This is a very dangerous form of artificial stupidity and makes us consider the power of what we are really creating. Although the behaviour of these systems could be modified if we modify the goal, it is alarming that in reality, we could deploy these kinds of systems into the world.

This is my fear of AI, not really intelligence itself but some bad human use case of intelligence or some poorly defined goal achieving system like DeepRL that is given too many resources. One can see the true nature of what it could be like if a poorly designed system is given too much execution power if one watches the Phillip K. Dick's film adaptation Electric Dreams episode, Autofac.

We must have a very objective view of artificial systems, rather than shout and blab about superintelligence we should see the effect of giving a poorly designed algorithm too much execution power.

The fear of AI brings to mind the fear of Nuclear annihilation. Nuclear power, the power of the Atom is just what it is power. Nuclear power could be harnessed to provide us with abundant electricity at some environmental cost and risks and Nuclear power could be harnessed to give man great powers of devastation if its goals deviate from healthy purposes.

The fear is not about nuclear power itself, that is just a tool for achieving ends. The fear is about the humans who would want to use this power to achieve destructive ends.

So is this era of AI as a tool in which we are currently in. Although many people with overactive imaginations will like to imagine that AI algorithms have something of an intention of their own, it is up to clear thinking people to clear out all the misconception. That is one goal of this work.

When I see some news article like AI found to be biased! The public will interpret it like: The overlord system that is winning everybody at every conceivable game is biased towards a certain group of people, oh I should be scared because if this system does not like me it might kill me in the future. The right interpretation is: “some engineer did not really assess their data properly before training their neural network (or any other kind of algorithm) and thus it is biased towards certain groups of people.

When we see the performance of Deep RL game playing systems, we are usually impressed about their performance and this can really fire up our imaginations but we should quickly get down to reality as soon as possible. The way Deep RL systems learn is not practical in the real world without doing millions of rounds of making mistakes in simulation. With time they will get very good at doing this because some smart programmer will figure out how to do simulations of the real world effectively making it possible for DeepRL systems to be deployed in the world.

But the fear in this systems is not that they have some kind of self-will of their own and like many people say will begin to formulate goals of their own. This is bad thinking, any goal a DeepRL system is executing will be what it has been given to it by humans at some time.

A DeepRL system with poorly defined goals will end up doing undesirable things in the world and this is not an effect of intelligence because it was done through sophisticated programming. It is as a result of poor design and not understanding the target problem properly.

And like Stephen Wolfram has been proposing in his talks about AI ethics ( https://blog.stephenwolfram.com/2016/10/a-short-talk-on-ai-ethics/)  we need some kind of language in which we can specify goals explicitly, no different from the way algorithms are specified to run on a computer. We need a language where we could say things that mean only one thing, a language robust enough in which we can specify something as serious as a legal contract. With such a language and I think the Wolfram Language is the foundation of such, we can specify goals to AI system that are unambiguous and indeed can be executed on a computer and have only one kind of output, meaning only one kind of thing at a time.

If there is a solid computational language like that then we could easily state the exact kind of goals we expect of a DeepRL system to achieve and avoid the kinds of scenarios that Lex Fridman always talks about in his lectures where rather than finishing the game the AI gets stuck at some point where points are abundant.

You should be afraid of human intentions when it comes to AI and not the AI algorithm itself which is simply a tool.

Comments

Popular posts from this blog

Software of the future

This powerful CEO works remotely, visits the office only a few times a year

At the edge of a cliff - Quantum Computing