Hello is not just a greeting. The moment you open this page, we are connected. No mater who you are or where you are, it is my honor to be known by you from 6 billion people on this planet. I don't believe you open the page randomly, cause there are billions of pages been viewed and visited every second, but why me, why this one, why this moment. So just keep reading, regard this as some kind of asynchronous communication. If you want to tell me something, hit the connect button and leave me a message, then we will be strangers no more!
Welcome again to my home in this internet world. The owner of this piece of land is currently a master student studying Robotics at University of Pennsylvania, whose ultimate academic goal is to build the 'brain' of robots! If robots only mean machines or mechanical devices, you might want to think it again. To me, robot is a life form who shares this world with us. It could be a piece of code, or a circuit board or even your smart phone. But this kind of life is so young and not yet well developed, which means it is not clever enough and always need human beings, its creators, to help them to finish their tasks. It is obviously not good enough!
Fortunately, lots of amazing achievements have been made in mathematics and statistics in the past few years, which recently popularize a field of study, Machine Learning. I believe machine learning gives us a key to eventually open the door for robots to be able to think and behave just like us -- human beings, or at least much smarter than those we already have. I can feel that moment is coming and I am so eager to evolve into it.
However, it is not a easy journey to build smarter robots. If we want to apply machine learning to robotics, we need to solve two major problems. The first one is where and how to get the data that robots need to learn from. All the successful exampleds of machine learning we have are the applications on the big data. When we train a robot, where the data come from? Please recall growth of a human being, parents and teachers spend at least 16 years to train a person to be mature. Thus the next question we need to ask is do robots need teachers? If the answer is yes, how should these teachers teach robots and how many years a robot need to learn before he can do work independently? Even if we find a way to answer all those questions above and successfully gather huge amount of data we can use, there will still be another open question about data themselves, in which format we can represent them so machines can use them.
The second major problem is what model or models we should use to teach or control robots. It is actually depends on the architecture of robots we design in the future. On high computational power platform, we may use some fancy but time consuming algorithm, like deep learning; on the other hand on low computational power platform we may need some fast or updatable algorithm, like SVM or perceptron. However is it the way we want to design robots, from a specific case to another specific case? Is there a way we can generalize the scheme? Machine learning people knows testing an observation or a group of observations is always much faster than training a model. Therefore, the next question should be asked is can robots share the same brain, which might be a network that contains thousands of machines. This kind of network will make training any kinds of model easier and faster, and only need to push back models to robots, which can be used by robots to make decisions. Cloud Robotics seems to give us a very promising solution.
Hopefuly, one day I will have the opportunities to work all those questions out and achieve something that will reshape this world permanently.