Down to Earth to look at the stars

It's been a year since we've been in contact with deep learning. Still remember last year at this time just fill out the graduate student's push-free, did not have to leave for four years has been some tired of Harbin, the heart is a bit frustrated. The only thing that may be fortunate is that at least you have escaped from the previously less interesting automation disciplines, and have great expectations for the direction of deep learning or artificial intelligence that will be learned in the future.

After entering the lab, the first group will report that I am responsible for introducing CNN to the whole group of students. At that time, I was also the first to hear the concept, only crazy learning the foundation of Evil. Open classes, papers, websites, blogs, everywhere is my learning resources, I seem to open a new world, or like a rustic of the first time a countryman to the big city. From the neuron to the full connection to the convolution, from the former to the reverse propagation, from the loss function to various optimization methods, each is gradually conquered by me. However, after the implementation of their own code is really clear, in the group to let other people understand the whole may be considered to be really understanding. At this time, I can only be entered into deep learning this door.

After that, we began to systematically watch Wunda's deep learning and machine learning series, while reading some of the more detailed papers to understand how deep learning is applied to specific image processing problems, and on the other hand, graduation design is also followed by a senior to try to apply deep learning to the three-dimensional mesh denoising, it can be said that The theory, the frontier and the practice are not falling. It seems that everything is going very well. Other people think so, the teacher think, I once felt so.

However, after knowing most of the classical neural frames, trying out a variety of optimization methods, and learning more and more assistant skills, I gradually felt that the black box did not look so dazzling at the time. Lack of perfect theoretical guidance, may let you go to the wrong direction, blindly pursuit of performance, but also let you encounter problems, experience first. Endless network variants and a variety of learning frameworks, not real AI, and the influx of countless peers will only weaken your true value, so that you in the talent market worthless. This is not what I want.

It should be said that this time the deep learning of the popularity, to a large extent, thanks to the massive data of the easy access and computing power greatly improved. All said, standing in the air, pigs can fly, but we all ignore the wind stopped after the pig fell how miserable. In such a person's mouth is not close to deep learning, we should be quiet to do something really conducive to our future development, the difference in competition can really go further!

As a non-trained, solid data structure and algorithmic programming ability is indispensable; in the field of image processing, good linear algebra, probability theory, matrix analysis and optimization are the key to winning; artificial intelligence may be more likely to be based on unsupervised learning and reinforcement learning; of course, Traditional machine learning and pattern recognition can be of great value at all times, just as all classical theories stand the test of the Times. These, perhaps, is my next graduate career should be the direction of real efforts, the right path coupled with continuous deep-rooted, I believe there will be fruitful output.

Recently read a paragraph, the effect is that if you do one thing every time there is a 51% probability of success, and can be repeated indefinitely, then this thing must be worth doing. I hope all of you can find these things worth doing soon, and then, continue to adhere to the ground, one day, you will find the sky is in sight!

For more highlights, please follow "seniusen"!

---------------------This article from Johnson csdn Blog, full-text address please click: 82914200?utm_source=copy

Down to Earth to look at the stars