Having been working at TP-LINK company for1year, I start to realize that what an important role artificial intelligence (AI)play in industry. The more I participate in the production of InternetProtocol Camera (IPC), the more strongly I feel the great demand of IPCrelevant research, like computer vision, multimedia, big data, and machinelearning, and so on.
For a long time, I have great doubt thathow a machine learning algorithm without 100% accuracy can be applied to themanufacture. I think every problem must be solved with100% sure, like 1 plus 1must be 2, but the machine learning solution will unavoidably result inuncertainty and the training set cannot cover the whole world.
However, as I step into the production ofIPC, I find that almost every link of production chain and every component ofIPC are filled with uncertainty. Even with the same IPC and the same scene, wecannot get two same pictures at nearly the same time. Given this uncertainty,we can imagine how difficult it can be to find a 100% accuracy solutiondirectly with manually learning from data. When I apply a face detectionmachine learning framework to the dust spot detection problem and get muchbetter performance than traditional method, the first time I see what anamazing idea to equip the machine with intelligence and let the machine learningfrom data!
On the other hand, when I have the StanfordMachine Learning class, at the Learning Theory chapter, I clear up all thedoubt about machine learning. Through this chapter, I realize that even thoughin machine learning algorithm the training set cannot cover the whole worlddata and the training error is not same with generalization error, the uniformconvergence of the gap between the training error and generalization error cangive guarantees on the generalization error, which is the theoretical supportof machine learning’s effectiveness.
Having recognized these facts about IPCproduction and machine learning, I’m more certain about my future career as ascientist of computer vision, which is a part of the artificial intelligencefield and closely related to machine learning, image processing, and big dataand so on.
However, my experience about computervision began at the third year as an undergraduate student, when I was anintern at multimedia lab of Peking University. It’s thefirst time I participated in the development of vision system which is used todetect the uploaded illegal videos. I was working on the subtitles extractionand acceleration of key-frame extraction, also fixed some bugs on softwaredesign. I participated in their group discussion and found that algorithm isthe kernel of the computer vision system. What an interesting experience if Ican design my machine learning model and optimization algorithm, so as tomanipulate the performance of system.
However, I realized that what I had learnedas an undergraduate is far from enough to do research. So, I gave up theopportunity to pursue the master degree of software engineering in BeihangUniversity on recommendation and joined the State Key Laboratory of IntelligentTechnology and Systems of Tsinghua University through the national postgraduateentrance examination.
The first year at Tsinghua University, accordingto my advisor’s advice, I got all the credits that school provided by selectingall computer vision relevant course. On the numerical analysis course, the firsttime I knew the gradient descent algorithm. However, I didn’t recognize that it’sthe most famous used method in solving optimization problems until I learnedthe machine learning course and understood some machine learning models’optimization methods. When I learned the EM (Expectation-maximization)algorithm, I was so deeply motivated by its perfect use of Jensen’s inequalityin solving the optimization problem with latent variables that want to sharethe excitement with everyone. I do homework of every course very carefully andgot excellent scores.
At the same time, I worked with my advisoron research. However, at the first time, the task is relatively simple, such asextracting image feature and completing some algorithms of video summary.Having been familiar with the video summary procedure, I tried to apply thesparse coding method to the topical object discovery of commercial video, andpublished the paper on the journal of Communications in Computer andInformation Science. On the other hand, I worked out the Gibbs Sampling methodto solve the optimization problem of soft Latent Dirichlet Allocation model andapply patent successfully. In addition, I published three other second-authorpapers cooperating with my advisor.