Machine learning, from deep neural networks to deep forests

Machine learning, from deep neural networks to deep forests

"Now we are in an era of big data, but I think everyone is very clear now that this big data does not mean really great value. To get the value in the data, we must conduct effective data analysis. Today, we have to use computer to analyze data, we must have machine learning." Professor Zhou Zhihua has repeatedly delivered his speech to explain his cutting-edge thinking about machine learning.

Regarding the question of what deep learning is, Professor Zhou Zhihua explained to you through a case. This example describes a well-known society called “Siam”, whose flagship newspaper, Siam News, talks about deep learning in machine learning as a “child”. Professor Zhou further talked about "depth": Physiology can get a basic inspiration, that is, a cell can get a lot of signals, this phenomenon was summed up by two scholars in 1943 into a simple mathematical formula: several sums, Then subtract a number and use the function to process it. This formula is still in use today. What is the mystery of seeing such a complex deep learning system today? Professor Zhou believes that the so-called neural network, deep neural network, the most basic unit is such a simple function.

Professor Zhou also pointed out in the paper that deep neural networks have made remarkable progress in the field of machine learning in the past decade. Multi-layer characterization is considered to be a key element of deep neural networks. By constructing a hierarchical or “deep” structure, the model can learn good characterization from raw data in a supervised or unsupervised environment, which is considered to be its success. The key factor. Successful application areas include computer vision, speech recognition, natural language processing, and more.

At the same time, although non-micro models such as the gradient boosting decision tree (GBDT) are the main methods for modeling discrete or tabular data, they are difficult to integrate this characterization learning ability. The multi-layer GBDT forest (mGBDT) proposed by Zhou Zhihua team explores the ability to learn hierarchical representation by stacking multiple regression GBDT layers as building blocks. The model can be jointly trained using variants of inter-layer target propagation without deriving back propagation. And microfeasibility, experimentation and visualization demonstrate the effectiveness of the model in terms of performance and characterization of learning ability.

The development of artificial intelligence industry, what is really needed is talent

In addition to studying the frontier topics of machine learning, Professor Zhou Zhihua is also very concerned about the industrial development of artificial intelligence. Although artificial intelligence technology has been widely used in many industries, in the view of Professor Zhou, the main application area of ??artificial intelligence technology is the Internet industry. He believes that machine learning technology provides a way to use data to solve a large amount of data accumulated by the "human" subject in all walks of life. "Future artificial intelligence technology should enter all industries like electricity." Professor Zhou pointed out.

What are the core elements of the development of the artificial intelligence industry? Professor Zhou believes that what is really needed is talent.

Professor Zhou said that the artificial intelligence industry is an industry that is very personal and intelligent, especially compared with the general software industry. Artificial intelligence is the jewel of the software industry. In some ways, the intelligence of individual talents can achieve an algorithmic breakthrough, which will really improve productivity. The characteristics of the artificial intelligence industry determine the talent cultivation of artificial intelligence is particularly important. In the era of global competition for artificial intelligence talents, the hematopoietic ability of high-level artificial intelligence talents will lead to differences in the core competitiveness of the industry, and the artificial intelligence talent training base will directly promote the formation of talents in the artificial intelligence industry.

The forthcoming China Artificial Intelligence Conference (CCAI2018) has become a high-end cutting-edge communication platform that closely combines production, learning and research in the field of artificial intelligence in China. It has launched a wonderful speech by the top experts in the global AI field to help the country to cultivate. High-level artificial intelligence talent. Professor Zhou Zhihua will discuss the elements of deep learning in the main report of the conference, and believe that deep learning does not necessarily have to be realized through the neural network model. Professor Zhou Zhihua will introduce his preliminary exploration of deep forests, a deep learning model based on decision tree forest rather than neural network.