Category Archives: Learning

Why AlphaGo is not AI

President at Novaquark

What is AI and what is not AI is, to some extent, a matter of definition. There is no denying that AlphaGo and similar deep learning approaches have managed to solve quite hard computational problems in the last years. But is it going to get us to AI, in the sense of a fully general intelligent machine, or “AGI”? Not quite, and here is why.

One of the key issues when building an artificial general intelligence is that it will have to make sense of the world for itself, to develop its own, internal meaning for everything it will encounter, hear, say and do. Failing to do this, you end up with today’s AI programs where all the meaning is actually provided by the designer of the application: the AI basically doesn’t understand what is going on and has a narrow domain of expertize.

The problem of meaning is perhaps the most fundamental problem of AI and has still not been solved today. One of the first to express it was Harnad, in his 1990 paper about “The Symbol Grounding Problem”. Even if you don’t believe we are explicitly manipulating symbols, which is indeed questionable, the problem remains: the grounding of whatever representation exists inside the system into the real world outside.

To be more specific, the problem of meaning leads us to four sub-problems:

  1. How do you structure the information the agent (human or AI) is receiving from the world?
  2. How do you link this structured information to the world, or, taking the above definition, how do you build “meaning” for the agent?
  3. How do you synchronize this meaning with other agents? (otherwise, there is no communication possible and you get an incomprehensible isolated form of intelligence)
  4. Why does the agent do something at all rather than nothing? How to set all this into motion?

The first problem, about structuring information, is very well addressed by deep learning and similar unsupervised learning algorithms, used for example in the AlphaGo program. We have made tremendous progress in this area, in part because of the recent gain in computing power and the use of GPU (Graphical Processing Units) which are especially good at parallelizing information processing. What these algorithms do is taking a signal that is extremely redundant and expressed in a high dimension space, and reduce it to a low dimensionality signal, minimizing the loss of information in the process. In other words, it “captures” what is important in the signal, from an information processing point of view.

The second problem, about linking information to the real world, or creating “meaning”, is fundamentally tied to robotics. Because you need a body to interact with the world, and you need to interact with the world to build this link. That’s why I often say that there is no AI without robotics (while there can be pretty good robotics without AI, but that’s another story). This realization is often called the “embodiment problem” and most researchers in AI now agree that intelligence and embodiment are tightly coupled issues. Every different body has a different form of intelligence and you see that pretty clearly in the animal kingdom. It starts with simple things like making sense of your own body parts, and how you can control them to produce desired effects in the observed world around you, how you build your own notion of space, distance, color, etc. This has been studied extensively by researchers like Kevin O’Regan and his “sensorimotor theory”. It is just a first step however, because then you have to build up more and more abstract concepts, on top of those grounded sensorimotor structures. We are not quite there yet, but that’s the current status of research on that matter.

The third problem is fundamentally the question of the origin of culture. Some animals show some simple form of culture, even transgenerational acquired competencies, but it is very limited and only humans have reached the threshold of exponentially growing acquisition of knowledge that we call culture. Culture is the essential catalyst of intelligence and an AI without the capability to interact culturally would be nothing more than an academic curiosity. However, culture can not be hand coded into a machine, it must be the result of a learning process. The best way to start looking to try to understand this process is in developmental psychology, with the work of Piaget or Tomasello, studying how children acquire cultural competency. It gave birth to a new discipline in robotics called “developmental robotics”, which is taking the child as a model (as illustrated by the iCub robot, pictured above). It is also closely linked to the study of language learning, which is one of the topic on which I mostly worked as a researcher myself. The work of people like Luc Steels and many others have shown that we can see language acquisition as an evolutionary process: the agent creates new meanings by interacting with the world, use them to communicate with other agents, and select the most successful structures that help to communicate (that is, to achieve joint intentions, mostly). After hundreds of trial and error, just like with biological evolution, the system evolves the best meaning and their syntactic/grammatical translation. This process has been tested experimentally and shows striking resemblances with how natural languages evolve and grow. Interestingly, it accounts for instantaneous learning, when a concept is acquired in one shot, something that heavily statistical models like deep learning are not capable to explain. Several research labs are now trying to go further into acquiring grammar, gestures and more complex cultural conventions by this mean, in particular the AI Lab that I founded at Aldebaran.

Finally, the fourth problem deals with what is called “intrinsic motivation”. Why does the agent do anything at all, rather than nothing. Survival requirements are not enough to explain human behavior. Even perfectly fed and secure, humans don’t just sit idle until hunger comes back. There is more, they explore, they try, and it seems to be some kind of intrinsic curiosity. Researchers like Pierre-Yves Oudeyer have shown that simple mathematical formulations of curiosity, as an expression of the tendency of the agent to maximize its rate of learning, are enough to account for incredibly complex and surprising behaviors (see, the Playground experiment done at Sony CSL). It seems that something of the sort is needed inside the system to drive its desire to go through the previous three steps: structure the information of the world, connect it to its body and create meaning, and then select the most communicationally efficient one to create a joint culture that enables cooperation. This is, in my view, the program of AGI.

Again, the advances of deep learning and the recent success of this kind of AI at the game of Go are very good news because lots of very useful applications can be imagined from there to help medical research, the industry, progress in environment preservation and many other issues. But this is only one part of the problem, as I have tried to show here. I don’t believe deep learning is the silver bullet that will get us to true AI, in the sense of a machine that is capable to learn to live in the world, interact naturally with us, understand deeply the complexity of our emotions, cultural biases and ultimately help us to make a better world.

This article originally appeared on LinkedIn

25 Ways To Ask Your Kids “So how was school today?” Without Asking them “So how was school today?”

NOTE:  This Post is now an APP that is available for purchase for 99 cents.  Every weekday all year round it will provide you with a different question to ask your kids…and none of the questions are “So how was school today?”  :)  You can find the APP HERE

This year Simon is in 4th grade and Grace is in 1st grade and I find myself asking them every day after school, “So how was school today?”.

And everyday I get an answer like “fine” or “good” which doesn’t tell me a whole lot.

AND I WANT TO KNOW A WHOLE LOT!!!!

Or at get at least a full sentence.  So the other night I sat down and made a list of more engaging questions to ask about school.  They aren’t perfect…but I do get at least complete sentences…and some have lead to interesting conversations…and hilarious answers…and a few insights into how my kids think and feel about their time at school.

25 Ways to ask your kids how was school

#1.  What was the best thing that happened at school today?  (What was the worst thing that happened at school today?)

#2.  Tell me something that made you laugh today.

#3.  If you could choose who would you like to sit by in class?  (Who would you NOT want to sit by in class?  Why?)

#4.  Where is the coolest place at the school?

#5.  Tell me a weird word that you heard today.  (Or something weird that someone said.)

#6.  If I called your teacher tonight what would she tell me about you?

#7.  How did you help somebody today?

#8.  How did somebody help you today?

#9.  Tell me one thing that you learned today.

#10.  When were you the happiest today?

#11.  When were you bored today?

#12.  If an alien spaceship came to your class and beamed up someone who would you want them to take?

#13.  Who would you like to play with at recess that you’ve never played with before?

#14.  Tell me something good that happened today.

#15.  What word did your teacher say most today?

#16.  What do you think you should do/learn more of at school?

#17.  What do you think you should do/learn less of at school?

#18.  Who in your class do you think you could be nicer to?

#19.  Where do you play the most at recess?

#20.  Who is the funniest person in your class?  Why is he/she so funny?

#21.  What was your favorite part of lunch?

#22.  If you got to be the teacher tomorrow what would you do?

#23.  Is there anyone in your class that needs a time out?

#24.  If you could switch seats with anyone in the class who would you trade with?  Why?

#25.  Tell me about three different times you used your pencil today at school.

*****

So far…my favorite answers have come from questions #12. #15, and #21.

I actually love questions like the “alien” one (#12).  They give kids a non-threatening way to say who they would rather not have in their class, and open the door for you to have a discussion to ask why, potentially uncovering issues you didn’t know about before.

And the answers we get are sometimes really surprising. When I asked question #3, I discovered that one of my children didn’t want to sit by a best friend in class anymore — not out of a desire to be mean or bully, but in the hope they’d get the chance to work with other people.

Sometimes we just need to figure out the right kinds of questions to ask our children….some questions may work better for some kids than others.  That’s how it is with my own children.  But I want to know what is going on in their lives and how I can help them.  So….I will continue to ask…and ask…and ask…

And, as my kids get older I know that I am going to have to work harder and harder to stay engaged with them…but I know its going to be worth the work…

-liZ

NOTE:  Because these questions were geared for elementary school aged children I have also made another list for teen age children that you can find here:  25 Ways to Ask Your Teen “How Was School Today?” without asking them “How Was School Today?”

****Also, to get a printable version of this list just click here:  25 Ways To Ask Your Kids So How Was School Today Printable

NOTE:  There is also a companion post for this post that you can find here:  17 Tips for Communicating with Kids

This article originally appeared on Simple Simon and Company