GEB notes

Intelligence depends crucially on the ability to create high-level descriptions of complex arrays, such as chess boards, television screens, printed pages, or paintings.

Talking with a Physics Researcher

Just had a very happy Christmas trip to Colorado. Skiing, hiking, and Colorado is quite a nice place. And, next summer, CVPR is going to be hold there.

During the whole break, I've been talking with my uncle, who is a physics researcher in University of Colorado, Boulder. Sometimes it is hard for a Computer Science guy to talk efficiently with a physics researcher. Apparently there are several things we share similiar opinions, and, expected, a lot of different ones.

First of all, when we talking about Artificial Intelligence, even he knows this area has stucked in a bottle neck for a long time. Like a scientist, he did asked questions like, What's the biggest improvement in the last 20 years of this area? Seriously, I have no idea on how to answer it. Maybe 20 years ago, the face recognition can reach 80% accuracy while now we can get 98% accuracy. Then he said: It is all because the improvement of computation ability, or hardware improvement. At the first thought, I want to argue with him that there are a lot of software improvement but later on, after a while of thinking, I have to admit that most of the improvements come from hardware. 20 years ago, the image resolution is low and maybe it was still black and white pictures, and 20 years after, we can get high resolution images to process on. 20 years ago we cannot extract complex features in a reasonable time and now we can easily extract very complex, such as SIFT, features to process on.

Then we talked about atom computation, which he thinks is the future of computation (he studies atoms or something like that...). He think the whole bottle neck of CS is the computation ability. But, atom computation is still in theory level and it is really difficult to do real experiments to make it work ( I totally have no idea why it is difficult, but apparently it is). I said something about DNA computation which I heard from a lecture, and apparently he did not believe it. He argued that someone thinks everything in the pysical world is computation, but something like DNA manipulation is just one of normal natural biology processes.

When we talking about my research area, such as vision and language, is it real intelligence? he asked. What is real intelligence? I don't know...  Then I said that some researchers in Computer Science have already abandoned those fancy stories of old school AI and only regard machine as a kind of large scale data analyser, and I talked something about Machine Learning. Apparently, a physics researcher buys the idea of Machine Learning or Statistical Learning.

After all, he hold a strong belief that Computer Science is a part of Physics ( or every science is a part of Physics), while I argued that if you regard a physics phenomenon with input, output and central process, it is a kind of Computer...

Holidays end, new semester starts~

First Semester is Over

After Machine Learning exam and 4 hours final grading, the first PhD semester is officially over.

A discussion about what is the valuable research went on in our lab two days ago. Machine Learning is quite hot topic for the last 10 years, and it will be. The whole magic about ML is that those statistic based algorithm can build models to predict output from input, without a deep understanding of the actual mechanism between input features and output, for example, when people use SVM to learn where ppl look (a preliminary step of attention mechanism), actually still we do not have a deep understanding of the attention mechanism, even the SVM do provide reasonable and accurate prediction. In a word, ML treat the underlying mechanism as a Black Box.

On the other side of the story, traditional AI researchers working on understanding the basic principles of those mechanisms, such as how do ppl analyse video, how do they learn from the environment, and how do they build their knowledge system. In the last lecture of Computational Linguistic class, Professor Philip Resnik 

(he is a really good teacher and going to take his Computational Linguistic II next semester) talked about an interesting concept in AI: AI-complete. For my understanding, AI-complete problems is a group of AI problem which can not be solve without a whole active knowledge system. If we can build an active knowledge system to solve one problem of them, others can be solved in similar way. Such as human recognition, a very basic research topic in Computer Vision, without a knowledge system, we can only recognize a group of shapes and textures looks really like a human, but who is he or she, what he or she have done and what should I do? All of those questions need an active knowledge system to answer.

Time flies, another Christmas and New Year is coming. I'll merry you Happy Christmas and Happy New Year!

Amazon Mechanical Turk

To evaluate subjective experimental result on Amazon Mechanical Turk is a lot of fun. We spent just 1 dollar, and get 80 responds.

More interesting, those workers ( or they might be a graduate student like you and me) gave us very serious and special comments. I collected them and posted here.

In one of our test image, we accidentally recognize a girl as a boy, and the worker said: Might be a girl, not a boy.

The same thing happened when we recognize a cow as a sheep: I believe it is a cow, not a sheep.

Well, most of the replies are like: Thank you; Good Hit; Nice to do this hit...

And the result we collected are pretty reasonable and useful for our Computational Linguistic final project.

Well, if you want to try the feeling of being a King, just spend several dollars to Amazon Mechanical Turk. 

And after did 7 missions, I also earned 1.** dollars...

Repeat other's work

The very strange thing in computer science, which is quite different from other research domains, is that seldom of researchers are working on repeating and validating other's work.

Two days ago, at the Poster session of the final Machine Learning projects, a classmate of us present nothing but a work repeat a state-of-the-art paper, which claims that they can achieve 90% accuracy in gender prediction on blog data. The classmate reported that he tried almost his best using the method proposed by this paper and the best he can get in a well-known dataset is a little bit more than 80%. Considering the difficulty of gender prediction, he did not doubt but was very curious about how the published paper can get more than 90% accuracy.

I talked with him at his poster later. I think most of us have the same concern. He said, if a new researcher come into a specific area and find that previously someone have already get such a high accuracy, he probably won't dig into this problem later.

A lot of interesting work are presented in the final poster session.  Wiki documents recommendation system,  predicting location based on your friends location on social networks and so on. And our group got an frog which can dance and sing as the best poster prize ^^

Knowledge System

Artificial Intelligence is all about Knowledge System.

Linguistic guys claim they are building some kind of Knowledge System based on language. Vision guys are always talking about semantic analysis. And ML guys usually talk about how to extract knowledge from data.

Where is the knowledge system.

How to build a knowledge system.

And how to use a knowledge system?