Marek Rosa reflects on the differences between evolution and intelligence and the implications for the development of artificial intelligence.
Read about it on his blog.
This talk has been edited for clarity and length.
Transcript:
My definition of intelligence is that it’s a tool that we as people use for achieving goals – it can be any kind of goal. I think about it as a tool that can be used to represent the real world, which is very complex, but pared down into a simplified, abstract model where only the relevant parts of the real environment are modeled. Those parts that are not important are not modeled.
These are some of my thoughts on how I came to this definition and why I personally like it. When I think about evolution as an optimization solution or algorithm, it’s not just about trying to find some solutions for something, but it’s trying to find solutions in the real world.
Take nest-building as an example. Birds build nests to protect themselves but this instinctive activity didn’t come about within one generation or one lifespan. Humans, on the other hand, think about how to cover their bodies when sleeping. There is a process of thinking and planning involved.
Birds don’t have intelligence so evolution optimized this problem by giving birth to new kinds of birds that had a tendency or urge to gather pieces of branches and put them somewhere. Basically by this random evolutionary process, there was a natural selection pressure which led to birds building nests.
The thing is that this took thousands or millions of years because evolution was trying to find this solution in the real world, which is extremely complex and has many parameters. You cannot use as many tricks for optimizing the search process as we can do with intelligence.
Evolution cannot plan, use abstraction, or test different hypotheses and then come back and use only the best hypothesis. This is the moment where intelligence comes in. It’s almost the same thing as evolution, except you transfer the problem to a simplified representation of the real world. Then you look at how to find patterns or how to use other tricks to find a solution or solve a goal in the simplified environment.
When I think about attention – what it actually is, in this simplified mental model you want to represent only those parts of the environment that are relevant for your goals. This is where attention comes in. If, for example, I feel that there is something I need to know more information about because my mental representation of this part of the world is not sufficient, or I think that there can be a benefit if I know more, I can either look there and receive new data and get a much more detailed representation of that part, or I can read about it, which is again just another kind of pattern, but the principle is the same.
Because intelligence works in a universe that has limited resources, you also want intelligence to work as optimally as possible, to use these scarce and limited resources as optimally as possible. You also don’t want to put attention to things that are absolutely irrelevant to your solution because this mental model that you have in your head always has some limits – you cannot have a bigger brain, time is a limited resource, and so on.
This is the similarity between evolution and intelligence. You optimize something and you use limited resources. When I compare intelligence and evolution they are very similar, except one operates in simplified mental representations and the other operates in a real, complex environment.
But the goal is the same. It’s just that you use different tools to achieve something.
About this representation or simplified model, I don’t think it’s a memory or pure memory of your previous experiences. I actually think it’s your previous experiences together with your current understanding of the world, which changes all the time. It’s your long term memory connected with your short term memory playing on the same battlefield. Being together, one helps the other and so on, but I wouldn’t really split them.
In my opinion, it’s not a memory or knowledge, but a representation of this complex, real world environment in your head. You look at the world around you and basically try to put as much as possible – of course only the relevant things – in your head, run these hypotheses, plan things, choose different things, decide if you want to put your attention on this thing or that thing so your representation can get richer in that area.
The best thing when compared to “dumb” evolution is that you can do this much faster than in evolution. Doing this in evolution would take a million years just like in the example with birds.
When I think about this representation, I think patterns are a really important part of the representation. Basically you, or a process of intelligence in the brain looks at this representation and tries to find patterns in various things. I can say as a joke that the more patterns you know, the bigger a person you are. You have more choices for making decisions and for learning new things.
Patterns are these little secrets. If you don’t know a pattern it is there in the environment and exists, you just don’t know it. You cannot connect those few things together.
Once you connect them together, you know that pattern and you can use it. This is one of the biggest miracles, in my opinion – finding patterns by yourself, inventing them by being creative, or copying or stealing them. That’s how you can grow knowledge in your brain and make a representation better and use it better.
It’s probably important to define a pattern. A good definition is that a pattern is something that happens in a regular and repeated way. A pattern can be an image where you see a pattern, but it can also be something that happens in time as a kind of sequence.
Patterns and sequences are very similar in my opinion. One more point regarding the uncertainty of the environment of the real world: the world itself is very uncertain. Sometimes we don’t know why something happens. The other thing is that sometimes you want to model certain aspects of the environment in this mental representation with probability because you don’t know for sure if these things really happen.
Sometimes you also want to model it with this probability because you don’t want to go into much detail. It’s like, there is some probability that I know this, this is true, and it’s ok for me – it’s sufficient to achieve my goals. When I think about this representation – and we are working on different models in our groups – I think about it as a probabilistic graph or graph of probabilities. It can grow in a certain direction where you want to get more details, specialized information, representation about that thing, or it can be really high-level, where you have just an abstract representation of certain things.
We can model this probabilistic graph, these ideas, with probabilistic languages, different neural networks, or Bayesian models. I don’t know which will be the winning strategy, but I am quite sure that it will be something with probability and hierarchy together because then you can get abstractions and all of these things.
I should also mention reasoning, which should be running on top of this representation, and looking for inference, deduction, induction – basically trying to find the correlations in the representation, trying to find new patterns, and also trying to learn casualties — something in your representation was caused by something else.
When you are searching for this, you make your representation much richer, and therefore you can find more patterns or useful patterns. The ultimate goal of intelligence is to actually use it for something.
If you have a lot of patterns and a lot of knowledge about causality, you can then use it for something. This is what evolution cannot do. Evolution doesn’t see the direct correlation between one object and another. With the abstraction that you have within your intelligence, you can actually see this thing.
In other words, for intelligence, you don’t actually need to try those things to move forward.
On comparing evolution and intelligence. Can evolutionary principles be used in artificial intelligence development?
I think it can. There are many evolutionary algorithms and optimization strategies.
We are using them (even now we have a genetic programming thing), but I think that even the evolutionary principles are “dumb” compared to intelligence. With intelligence you can look into the future and choose what you will or won’t abstract. Intelligence does this.
For example, think about a normal problem we have here: we have some code for our games that we want to optimize. Will we use evolutionary algorithms to optimize that piece of code, or will we use a person, a programmer, with his intelligence and experience and knowledge, to optimize that code?
I’ll always use that person, because it will just be faster. He will think differently than evolution would think. I think evolutionary principles can be used for some things, but I think that the optimization strategy of evolution is worse than intelligence. But to get there, to get to a high performing AI or at least to human-level AI, can we use some principles of human development stages from childhood to adulthood?
Yes, this is what we call gradual, guided learning. But again this is just a trick to speed up the learning process. We, as humans, have already gone down this path. I can say that our civilization has already searched through this space of possibilities and found a way that works. We don’t need our AI to repeat this search. Even with a good computer it would probably take millions of years.
Here AI can learn from what we have already learned, or what civilization already learned. We will have gradual and guided learning.
Guided means that somebody – a mentor, or us, or a predefined learning school – will guide the AI through the steps that we know are important to get some knowledge about the world, to basically have a really useful representation of the environment and not the non-useful things.
It’s gradual because we don’t want to throw everything at the AI – even patterns, highly complex patterns that the AI wouldn’t be able to understand. At the beginning we want to throw simple patterns at the AI to present it with a simple environment where it can actually find patterns in a way that it can grasp and move continuously forward.
For example, when you have a small child you don’t teach it how to program just by throwing programming books and a computer at the child. The child would not know what to do with it and would just be overwhelmed by these patterns.
But if you first teach it how to speak, write, and read, gradually you will get to patterns that represent programming things that can work. It’s actually much faster because you build up this pattern recognition hierarchy gradually.
Maybe one more example with gradual learning is that if you decide I’m going to learn the Chinese language – how to read and write in Chinese – you probably wouldn’t start by opening a Chinese book and trying to find the patterns in these Chinese symbols. That would probably take forever. You would need to encode. There would be hieroglyphs and such things.
What you can do is find a book, a Chinese for beginners thing, and you would start with understanding the absolute basics of Chinese, understand those patterns. You would move on to the next chapter, learn more complex patterns, move on to chapter 3, learn a bit more complex patterns. In this way, you will actually be able to understand Chinese.
If you have to decode all of these things, I think this would be, even for superhuman intelligence, a really hard job.
To sum it up, my final definition of intelligence would be that intelligence works in a simplified representation of the world where it searches for patterns that may help find a goal in an optimal way and to do it through a simplified mental representation because it’s cheaper and faster than doing it in the real world environment.
With intelligence you can have better structures and use different optimization strategies, better than evolution, to find solutions. Intelligence uses planning, abstraction, looking for correlations between events, how things relate to each other and so on.
Intelligence optimizes ways to achieve these goals and always looks for more efficient ways to use limited resources.
For the latest from our blog, sign up for our newsletter.
