Artificial intelligence has always been a topic of debate when it comes to how it performs and how it can perform better. One of the new ideas thought off is just how vast the internet is and how helpful it can be. The only problem is that the information available is in plain text, so extracting this information for analytical tests that can help humans see correlations between differnt things is difficult to do and time consuming, we all know that when it comes to computer science and machine learning, the time it takes for computation is very important and each computation must be efficient in its accuracy and speed. In general information extraction has always been more of a challenge in the realm of computer science and artificial intelligence just because there have not been any methods created to make this method work any better. Not until recently have there been students at the Massachusetts Institute of Technology along with researches that have brought the concept of machine learning to the next level. We have to realize that most machine learning work mostly by combining training examples that are provided to them by the human annotators. They just look for patterns that match keywords that are once again in memory, meaning they are truly limited in the information they can access because of their initial training. This new idea that has been brought up concerns machine learning and extends its capability by allowing the machine learning system to assign what they call a "confidence score" to each classification that the system makes. If this this score is too low then the system will automatically go searching the web for information that might contain the information that the system is trying to extract.
After this initial process the system then tries to extract all relevant data from the new texts that it found from its search query and combines this information with its information that it extracted in the initial attempt to find information. The cool thing that I am amazed by is the fact that all of this computing that is being done by the system is the result of machine learning, the system learns how to search, look at the accuracy of its search, and then be able to comprehend if the information is relevant to its initial task at hand, and then to be able to fuse the two together to complete its extraction. These researchers tested this system on many pieces of data that are extensive like mass school shootings and food contamination. The system was tasked with extracting things like names of shooters and schools, and in the case of food contamination, food type and type of contamination. The system was trained to respond to keywords and teach terms. In each trial the system was able to extract about 10 new documents from the web that related to the 300 documents that it was initially trained on. After seeing that this system actually works, these researchers compared their systems performance to that of conventional machine learning systems techniques. Basically one each test, the newly created system by the MIT students and researchers was able to out perform its older predecessors by almost 10 percent every time.
After this initial process the system then tries to extract all relevant data from the new texts that it found from its search query and combines this information with its information that it extracted in the initial attempt to find information. The cool thing that I am amazed by is the fact that all of this computing that is being done by the system is the result of machine learning, the system learns how to search, look at the accuracy of its search, and then be able to comprehend if the information is relevant to its initial task at hand, and then to be able to fuse the two together to complete its extraction. These researchers tested this system on many pieces of data that are extensive like mass school shootings and food contamination. The system was tasked with extracting things like names of shooters and schools, and in the case of food contamination, food type and type of contamination. The system was trained to respond to keywords and teach terms. In each trial the system was able to extract about 10 new documents from the web that related to the 300 documents that it was initially trained on. After seeing that this system actually works, these researchers compared their systems performance to that of conventional machine learning systems techniques. Basically one each test, the newly created system by the MIT students and researchers was able to out perform its older predecessors by almost 10 percent every time.
I find this amazing and it can only get better from here. These researchers are even thinking about adding millions and millions of articles to this system so that it can even perform on a wider scope. Machine learning just blows my mind away every time I hear about it. We are on the brink of allowing machines to do things that we can't do and faster. The best thing is we can train them on exactly what we want them to do so that they do only the task that they asked for. I can only imagine that this will get even better in the near future, and hopefully it will be for the better of science.
Reference Links:
http://image.slidesharecdn.com/machinelearningfordummies-140401055817-phpapp01/95/machine-learning-for-dummies-4-638.jpg?cb=1396332069
http://news.mit.edu/2016/artificial-intelligence-system-surfs-web-improve-performance-1110
http://news.mit.edu/sites/mit.edu.newsoffice/files/styles/news_article_image_top_slideshow/public/images/2016/MIT-Webaid-Learning_0.jpg?itok=Sli1Qec0
Really great article! I had no idea how AI's work and the depth to which they can be used. I think it is really important to understand the basics of AI's especially as we are programming our own games. Your article was really interesting because it took things a step further and making the AI system efficient for everyone. So cool!
ReplyDelete