7c0h

Articles tagged with "Research"

The Tapiz instruction-giving system

This article is the third of a series in which I explain what my research is about in (I hope) a simple and straightforward manner. For more details, feel free to check the Research section.

For my first research paper during my PhD, the basic idea was pretty simple. Imagine that, after recording several hours of people being guided around a room, I realize the following: everytime a player stood in front of a door, and someone told them "go straight", they walked through the door. So now I ask: if you are standing in front of a door, and I want you to walk through it, would it be enough for me to say "go straight", like before? My research team and I wanted to give this question an answer, so this is what we did.

We looked at our recorded data. Whenever we saw a player moving somewhere, we took notes about where the player was, where is the player now, and what was the instruction that convinced the player to move from one place to the other. We then created a big dictionary, where each entry reads "to move the player from point A to point B, say this". Quite smart, right?

The most important part about this idea is that we don't need to teach our computer how to understand language - in fact, when our system reads "turn right" in our dictionary, it has no idea about what "turn" or "right" mean. All our system cares about is that saying "turn right", for some strange reason, causes people to look to the right. This makes our system a lot simpler than other systems that try to understand everything.

Now, let's complicate things a bit: let's say I tell you "walk through the door to your left". You turn left, walk through the door, take 7 steps, give a full turn to look at the room, and then you wait for me to say something else. Which of those things you did because I told you, and which ones you did because you felt like it?

Since we didn't really know the answer, we tried two ideas: in the first case, we decided that everything you did was a reaction to our instruction (including the final turn), while in the second one we only considered the first action (turning left), and nothing else. As you can see, neither approach is truly correct: one is too short, and the other one is too long. But in research we like trying simple ideas first, and we decided to give these two a try.

Our results showed that the second approach works better, because if you advance just one step I can guide you to the next, but if you do too many things at once there's a chance you'll get confused and lost. Also, since our system is repeating what other humans said before, players thought the instructions were not too artificial.

Not bad for my first project, right?

What is the GIVE Challenge?

This article is the second of a series in which I explain what my research is about in (I hope) a simple and straightforward manner. For more details, feel free to check the Research section.

The GIVE Challenge is a competition started in the University of Saarland, created to collect data about human behavior. Since most of my research is based on that data, it's a good moment to explain what is it about.

We all know GPS by now - whenever we go by car somewhere new, we just type the direction and the GPS guides us. But have you ever thought about how hard it is to give instructions, like your GPS does? For instance, if we are in a roundabout and I say "take the third street to your right", does that mean I have to count all streets, or should I ignore wrong ways? And how much time do you need to react to my directions? These are important question, because they reveal a bit more about how humans act and think.

If we want answers, we need to collect data (reaction times, distance to other cars, misunderstandings, etc), and that data is very difficult to get. For our example, you would have to drive while wearing special glasses, a military GPS, and keep track of all the cars and pedestrians around you. So you might wonder, couldn't we make something simpler, but still useful? My adviser and other researchers asked themselves this exact same question in 2007, and that is how the GIVE Challenge was born.

In GIVE, a person sits in front of a computer, and they play a game. The game is pretty easy - all the person has to do is walk around a virtual house and press some buttons in a certain order. Just like a GPS, they receive instructions telling them where to go and what to do.

In the first variant of the GIVE Challenge, the instructions are given by a person using a computer in a different room. We then record all the information about how the player reacts to the instructions: if the instruction says "turn right", how much does the player turn? Do they just turn, or do they walk too? And how long does it take them? By recording every single movement of the player inside this game, we can answer questions like that.

There's also a second variant: we can write a program that guides the person inside the game, and see how good (or how bad) its instructions are. While a common GPS only cares about streets, our programs have a harder time: humans are not limited to just following streets like cars do, so the instructions are more complex. GIVE is a good way of testing how smart our computers are, and that's why we've been using it for many years now.

We've so far recorded over 340 hours of human movements, divided in 2500 games. Believe it or not this is not too much data, but it's a good start. We have extracted several interesting results from this data, some of which I talk about in future articles.

What are training and testing?

This article is the first of a series in which I explain what my research is about in (I hope) a simple and straightforward manner. For more details, feel free to check the Research section.

In research, we often want to teach computers how to do a new task, but that is difficult because computers are not too smart, and teaching them even a simple task takes a lot of work. So let's say I want my computer to tell me whether an e-mail is important or not. If I could teach my computer that, then it could show me important e-mails first and save me the trouble of sorting through them daily.

One way of teaching tasks to computers is by doing the job myself, and then make the computer repeat what I did. This is something scientists have been doing for a long time, and today we have a set of steps that every researcher should follow.

The first step is to collect as many e-mails as possible, both important and not. In science, such a big set of e-mails is called a corpus.

Now, just like you wouldn't know what kind of e-mails I consider important, neither does a computer. So the second step is to go through all those e-mails I collected, and mark which ones are important. I'll create two groups, one called "training" and another one called "testing". The first group will contain 4 out of 5 emails, picked at random, while the second group will have the remaining ones.

The third step, unsurprisingly called the training stage, requires the computer to analyze all the e-mails I put in the training group and decide what makes an e-mail important. We would expect our computer to understand, for instance, that since every e-mail containing the word "SALE" was marked as unimportant, then it might be a good idea to mark all e-mails with commercial offers as unimportant. This is by far the hardest step, and there are many ways in which I can influence how well the computer will learn.

The fourth and final step is to give our computer a test, to see whether it learned something useful or not. For this step, called the testing stage, I'll go through each e-mail from the testing group, show the computer the e-mail's text, and ask whether it's important or not. Then I compare the computer's answers with mine, and I'll use that result to decide how good (or how bad) my computer learned the task. If the results are not good enough I can always go back, change how are the e-mails analyzed, and try again. If the results are good, on the other hand, I can trust my program to sort my e-mail from now on.

This is pretty much half my daily work. Collecting enough data (e-mails) is either complicated, expensive, takes a lot of time, or all of that together. And remember I said there are several ways in which a computer might learn? We have to try some of those alternatives too.

Finally, training is usually very slow - in my last project, it took almost a week.

I usually dedicate that time to play Solitaire.