High-level Knowledge on the Web

Let computers “think” like a human being is probably the dream of almost every computer scientist. However, despite extraordinary capability in repetitive computation, computers totally lack the high-level knowledge that human possess. For instance, while a five-year old child can easily understand what the object is given a photo, the image understanding problem is still an open problem in the computer vision community. The same can also be true to many other problems, e.g., translation between two different human languages.

The root cause of this problem might be the essential difference between the Von Neumann architecture and the human brain architecture. Hence, even the most sophisticated algorithm can hardly mitigate this problem. Fortunately, we have other solutions besides derive algorithms from the scratch. One interesting approach is the human-based computation. That is, basically, cleverly design the system so that people are motivated to complete the task (one small chunk at a time) that are harder or impossible for machines to solve. Here I attach two videos that are great examples of human-based computation that covers two projects: reCAPTCHA and Duolingo.

Another promising approach is to utilize the information that already available on the web. The good part about contents on the web is that they are produced by human. So they contain the high-level human knowledge as long as they can be processed before utilized by computers. For instance, data-driven vision is a subfield of computer science, which use data on the web to solve computer vision problems that are otherwise hard or even impossible for computers. e.g., Hays et. al used web photos to fill the missing part (can also be use to re-combine) of given photos. Noted that it would be nearly impossible to achieve similar result by only using the low-level visual information such as pixel value, gradient, etc.

The same notion can also be applied to textual information. For example, teach the computer about common sense. Recently, Siri is popular because she has common sense so that she can act more like a human. e.g., if you say “good night” to her at morning, she will try to make fun with you.

Of course the common sense about “good night” is pretty simple and can be input manually. However, the common sense that human possess is so huge and subtle so that we still need the aid of human-based computation or utilizing the content on the web. There has already some great works been done at the MIT Media Lab in this direction.

Textual and visual input are the most important input to the human mind. Fortunately, it’s also the most prevailing kinds of content on the web. Before the day that computers can really “think” as human, how to cleverly transferred the human high-level knowledge to machines would be an exciting direction that can greatly improve our technology.

---