Tian Yuandong deep learning does not solve the theoretical problem (with 3 ICLR papers) – Sohu Techn coscoqd

Tian Yuandong: no problem solving theory of deep learning (report of 3 ICLR papers) – 1 new technology Sohu Zhi Yuan recommended source: know almost columns "Far East anecdote" Author: Tian Yuandong Zhi Yuan to start a new round of recruitment: COO, executive editor, senior editor, compiler, editor, director of operations and client manager Consulting Director, administrative assistant, 9 position overall opening. Resume: jobs@aiera HR WeChat: new Zhi Yuan COO and executive editor of the highest salary offer over 1 million; to provide the most complete training system, higher than the industry average wages and bonuses for the backbone of the staff. Join the new wisdom yuan, and artificial intelligence industry leaders to join hands to change the world. [introduction] new Zhiyuan Facebook artificial intelligence group researcher, go project director Tian Yuandong yesterday in the know almost writing a column, the three papers of ICLR2017, of which two papers (including the Doom AI Track1 application was first, a theoretical thesis). Tian Yuandong also elaborated his thoughts on the limitations of deep learning. ICLR2017 submission time finally cut off. The cast of three articles to go out, is my personal research since the largest number of single conference submission. Let the computer write their own code ICLR this meeting is open to the public, after uploading the article, anyone can see, there is no delay. Read the papers submitted, you will see the rapid iteration of this field, rare in the world. I was in the column "quick iteration of artificial intelligence," which referred to the speed of the field submission, wait until this time to look at the list of contributors, found himself may underestimate. Like "let the computer to write their code" the idea, last year also appeared very much, and is mainly to construct differential computer (such as nerve Turing machine, DeepMind differential neural computer) in the form of neural network optimization method to the end of the gradient descent, such as the society for sorting and other specific tasks. But this idea has several problems of relatively large, one is because the black box nature of the neural network, the computer can not explain the differential use of learning algorithm; second is to select the gradient descent method depends on the initial value, and the optimization process is slow; the third is the extension is not good, the number 60 can be sorted 100, the number of not, contrary to human intuition. And let the computer generated code does not have 1 and 3 of these two problems, the 2 can also be trained by the neural network to generate code to overcome, not through optimization, and use a forward spread on it. Second our submission is based on this idea, the input and output results of the algorithms for feature extraction, the classical framework into hierarchical convolutional neural networks in the literature to generate images, generate a two-dimensional map, each line is a line of code, or more precisely, is the probability distribution of the code. With a good distribution, it can help to find the correct procedures for heuristic search. The training data of the neural network, by a large number of random code.相关的主题文章:

« »

Comments closed.