 Hello everybody, I'm Harry Valpalla from the curious AI company and I'm going to talk to you about How the future AI will look like what's wrong with the AI now and how we can fix it and then what we can do with that So you must have seen many many different examples of artificial intelligence deep learning in particular has been a lot in the In the news It's been used for a machine vision machine translation games Whatever generating images all kinds of many very cool Applications and there's a lot of research. There's a lot of money in there a lot of hype But this AI that we now have Clearly isn't the kind of AI that we were dreaming of There are clear shortcomings of the AI that we now have In practice when you try to apply this this AI that we now have You need to have humans a lot of human work is needed and you need to have a lot of data That's the only way to get these deep learning systems Really working. That's the only way to get good results So now why is that? Why is it that we we need a human in the loop? What is human needed really for? I hope many of you have read Daniel Kahneman's great book thinking fast and slow So who has read this book? Who knows about this? Okay about say 30 percent of people Okay, so one of the things that he discusses in the book is System one thinking which is very fast and system to thinking which is slow the liberty thinking So we humans and also mammals in general we have these two different systems The first one is a system which can which looks very much like Deep learning it learns Learn slowly it needs a lot of data, but once it has learned it's really fast There's a little bit like automating some some actions or distilling knowledge This is basically what modern-day deep learning is doing in in most applications When you encounter new situations something you've never seen before your old Knowledge is no longer going to be enough in those cases humans resort to system to slow deliberate thinking Which means that we we can step back and we will notice that okay. There was something new. What's what's happening? It's increasingly it's Incredibly important to have that system in order to react to new things Without that we would be like reflex machines And reflex machines is how how deep learning looks like now Also if you think about it when you encounter new situations Any solution there is going to be creative So creativity is something that relies on system to it's something which humans are really good at and Computers not so much right So what we are now lacking in AI is this system to slow deliberate thinking That's not the only thing that we are lacking. There is also another shortcoming of Normal regular deep learning and that is that deep learning nowadays is based on learning associations between features The way deep learning systems perceive their world is is really through features You need a human to tell this system what objects are It will not be able to learn it all by itself It will not be able to learn about Objects if you only give it pixels videos so on so These two things You you can't learn quickly anything new and that this is system doesn't really understand objects. These are what is still Taking keeping AI back. That's why we need the human in the loop. That's why we need so much data You need to cover every possible case Because you can't rely on AI Coming up with some creative solution So now let's go a little bit deeper into technology. What is it that this magical system to does? For an engineer is important to understand how does this work? Otherwise, we are not going to be able to implement it So what I'm saying is that this system to Realize on planning and internal models So we humans have an internal model of how the world works we have learned that and when we Come up with a new situation. We are able to quickly adapt that model and Most importantly, we are using that model for planning. So we can make decisions based on updated knowledge. So for instance if I learn that the metro is now Down and I can't use the metro. I will be able to plan myself a new new route to to work for instance There are some cases where Neural networks have been combined with planning Very popular example is alpha go How many of you have heard about alpha go? The match against Lee settle. Okay, many of you you know about that. So it was a big thing Go the game of go is really difficult. It requires intuition system one thinking But alone. That's not enough. You also need planning and Alpha go is planning it is able to think about different Actions and it's able to learn from scratch. So you don't need to have a human at all To boost rapid because it's able to rely on its internal model of the game of go and Planning to come up with creative solutions from scratch Then its system one is able to learn to Automate this so it's a little bit like human experts do all the time once you have solved the problem enough times You will learn to automate it. It becomes a habit You you just create this intuition. So that's what alpha go is doing So I just said that a I can't do this and Alpha go does it so so what am I talking about? the trick is that What alpha go does not have and what doesn't work is that if you try to combine planning with learned Internal models that doesn't work. So the reason alpha go works is that the internal model is Hard-coded a human programming. So that's why you need a human in the loop there, too If you encounter a new situation and you are not able to learn Then your system 2 is very limited. You can't really cope with all all kinds of new situations So in order to really make good use of system 2 we will need to have learned internal models Neural networks can learn things, but it hasn't been possible to combine it with planning now We have technology which is doing this. So We are going to have this system 2 Capable AI shortly other people are working on that too. So that's what's coming up very shortly Another missing piece which I mentioned was this Neural networks which can understand objects and interactions neuroscientists have found coding in the brain which is able to Group features together into objects That's how our brain is able to perceive for instance different people in the audience or Understand different objects and we are able to learn that from just raw sensory input Nobody in our childhood has had to tell us what our cars and so on like Pointing out cars and drawing the boundaries. We figure out what the objects are Yes, we learned the names from our parents and so on but Language but still mostly we learn objects and their interactions just from observing raw data This is another technology which we are working on combining neural and symbolic technologies so These are the missing pieces I would argue something which we still lack The reason why we don't have The kind of science fiction artificial intelligence them that we were promised. This is why AI Isn't helping me do all my work But once we are done with this technology and as you heard from the intro introduction We are doing research. We don't have a product now But we do have an idea of what the product will be and that's an AI co-worker a System which is like a little bit like virtual assistant, but much more clever something which with which you can collaborate Just like with your human co-worker and Which will be able to learn and adapt quickly and which is clever enough that you can allow it to Work autonomously so that it can handle many of the tasks that you need to take care of yourself now So this is the kind of Future that we we see we will have these AI co-workers science fiction AI if you want if you will and they Will start collaborating not just with you and other people but also with other AI co-workers So that's where this internet of mind comes from That's how we are going to be able to augment human intelligence. We are still needed. Don't worry You're not going to be out of work These these systems we believe will be your your helpers and With with this kind of system you can increase the the potential of humans many fall so Welcome to the future. Thank you