Deep learning moves from being supervised to interacting

Source: http://tech.163.com/16/0427/07/BLL3TM9M00094P0U.html

Editor's note: 2016 is the 60 anniversary of Ai's birthday. April 22, the 2016 Global AI Technology Conference (GAITC) and AI 60 commemoration ceremony was held in Beijing National Convention Center, about 1600 experts, academics and industry members attended the conference.

The special report of the General Assembly is chaired by the Deputy Secretary-General of China AI Society and Dr. Kaiyu, founder and CEO of Horizon Robotics. Guests include the chairman of the Chinese Society of Artificial Intelligence, academician Li Deii of Chinese Academy of Engineering, IBM China Institute of Big Data and cognitive computing research director of the Middle Jiangsu, Baidu Deep learning researcher "outstanding Scientist" Xu Wei, Cambrian Technology founder and CEO Chen Tianxi, Beijing Cloud know-how Information Technology Co., Ltd. Chairman and CTO Liang and so on.

At the AI technology Conference, Dr. Yang Ming, co-founder and vice president of software at Horizon Robotics, delivered a speech explaining the new trends in deep learning development. Yang Ming said that since 2006, deep learning has exploded, mainly because of the huge amount of data used. The use of these big data makes some of the problems of this deep neural network no longer a problem.

Yang Ming that there are currently four new trends in deep learning, namely "MARS", the first is to learn how to remember (memory networks), and the second is to learn how to focus and trade-offs (attention model) and focus on the details that need attention The third one is the enhancement of learning (reinforcement learning), learning how to control the initiative, and a new trend in the fourth overall learning task structure, namely serialization (Sequentialization).

The following is a transcript of Yang Ming's speech:

Hello, I am Yang Ming. It is a great honor to have this opportunity to share with you some thoughts and summaries of our new trends in deep learning research, and we will abbreviate these new developments into a word Mars. This is a joint discussion with my colleague Dr. Huang.

Briefly, I joined the horizon last summer, and I was in charge of software engineering. Before that, I was in the Facebook AI lab in charge of face recognition algorithm research and back-end system development, and also worked in the NEC American laboratory and Xu Wei, learning a lot of things.

Definition of deep learning

Before we talk about new trends in deep learning, we should first clarify the definition of deep learning and its current state of development. Fortunately, the academic circle has a clearer understanding of the definition of deep learning. Deep learning refers to the expression or description of the data from the original data through constant learning and constant abstraction. So simply put, deep learning is learning its expression from raw data (learning representations). These raw data may be image data, may be speech, or text, and this expression is some concise digital expression. The key to deep learning is how to learn this expression. This expression is learned through multilayer nonlinear complex structures, which may be neural networks or other structures. The key is to hope that through end-to-end training, direct learning from data to expression.

If the origin of deep learning is to go back to the 1957, start with a very simple structural unit, the perception. Some input signals are weighted by weight, and a threshold value is compared to get output. Why is this the origin of deep learning? Because these weights are not pre-designed by rules, they are learned by training. The first "Perceptron" is the hardware design, which is the physical connection, which may be achieved by adjusting the resistor. At the time, the media predicted that it was a intelligent computer, and could soon learn to walk, speak, look, write, or even self-replicate or self-conscious. So after 60 years, it's time to look at the middle stage of the picture and writing, hoping to learn to replicate itself at least until 60.

The fall and rise of deep learning

Deep learning from the advent, in general, after the fall of two. Everyone was very optimistic at first, but soon found that there were some very simple problems that it could not solve. From 2006 onwards, under the impetus of Hinton, LeCun, Bengio, Ng and other professors, deep learning has been developed in an explosive way, with some breakthrough improvements in image recognition, speech recognition, semantic comprehension, and advertising recommendation. The latest development is the Alphago go competition this March, in a very intuitive way to make the community feel the progress of deep learning. We hope that in five years, deep learning technology can really be used in the daily lives of millions of households, so that each device can run deep learning modules.

In these ups and downs, the basic learning style and network structure of deep learning have no intrinsic change, or a multi-level artificial neural network structure. As seen in this picture, the input layer is some raw data and has annotations. No matter what you want to learn, as long as there is a function of evaluation error (cost function), evaluate the error of the neural network, then with this input and output, deep learning or deep neural network can be used as a black box to learn this goal. The artificial neural network is structurally the multi-layer neuron and the connection between them, combined into many layers. There may be an input and a target at the beginning, such as the person you want to recognize from the face image. This time the neural network certainly does not recognize it, because it has never been seen before. We will randomly set some values for the neural network to predict this recognition result, and the first output layer will almost certainly be a false recognition result. It doesn't matter, we take the error of the output layer slowly back, and modify the internal parameters of these neurons and the connection between them a little bit. With this little modification, the network learns a very complex function by using a lot of data. From the 80 's until now, this basic structure and learning algorithm has not changed in these 30 years.

Since 2006, there has been an explosion in deep learning, due to several reasons. The first is the use of huge amounts of data, the use of these big data so that some of the original deep neural network problems are no longer a problem (such as noise data sensitive, easy in a small data set performance is good, but can not be generalized to a large data set). The ability to learn with these big data requires a high level of parallel computing. Of course, there are algorithmic improvements, such as dropout, batch normalization, residual networks, etc., can avoid the cross-fitting gradient disappear these problems. But in essence this burst of deep learning is achieved through big data and computational power. Before saying that the neural network itself is like a black box, the structure setting is not very good guidance, this is still the current situation.

Why does deep learning get so much attention in the past few years? The key reason is that performance accuracy increases as data grows. Other methods of machine learning may increase in performance to a certain point as the data increases and becomes saturated. But so far we have not observed this in deep learning, which is probably one of its most noteworthy. At present, deep learning has also achieved many successes, such as how to do a good job of image classification problem. For a class 1000 image classification test, after about less than five years, the error rate from 25% to 3.5% of the level, has been more than the human recognition accuracy rate is higher. This is our current deep learning or deep neural network to achieve the main success point, that is learned how to identify, how to classify.

New trends in deep learning research

Back to our point, what is the new trend in deep learning research? We have summed up four directions. The first is to learn how to remember (memory networks), and the second is to learn how to focus and trade-offs (attention model), to focus on the details that need attention, and the third is to enhance learning (reinforcement learning), Learn how to control the initiative; The new trend in the fourth overall learning task structure is serialization (sequentialization). We take the initials and abbreviate it into Mars.

The first one is learning how to remember. Conventional feedforward neural networks have a feature: each time you input and output is a deterministic relationship, for a pair of images, whenever input into the neural network, we have a layer after layer of calculation will be a definite result, which is not related to the context. How do we bring the ability of memory into the neural network? The simplest idea is to add some state to the neural network so that it can remember something. Its output depends not only on its input, but also on its own state. This is one of the most basic ideas of recurrent neural networks. The output depends on its state, and we can also expand it into a sequence series structure, that is, the input of the current state includes not only the present input, but also the output of the previous moment, thus constituting a very deep network. This approach allows the neural network to remember some of the previous states. Then the output depends on the combination of these states and the current input. But this approach has a limitation: These memories will not last long and will soon be washed away by the data behind them. After the development of deep learning is long-term short-term memory, the concept of a memory cell, the unit added three doors, an input door, an output door, a forgotten door. Enter the gate to control whether your input affects the content of your memory. The output gate affects whether your memory is output and affects the future. The forgotten door is to see if your memory is self-renewing and keeps going. This way your memory is kept flexible, and the control of how memory is maintained by the doors themselves is learned by learning how to control these gates through different tasks. This short-term memory unit was proposed in 1999, and in recent years there have been some new improvements such as the gated recurrent Unit, which simplifies to only two doors, one is the update door, one resets the door, and controls whether the memory content can continue to be preserved.

These methods can actually save the memory a little longer, but it is still very limited. Some of the newer research methods present a concept of a neural Turning machine: There is a permanent memory module that has a control module to control how the memory is stored by the input to be read and converted into output. This control module can be implemented using a neural network. For example, for example, a sort of work, there are some random numbers, want to put it in sequential sequence. We need to design different sorting algorithms before, and this neuro-Turing idea is that we give these inputs and outputs and let the neural network learn how to sort the numbers by storing and removing them. In a sense, let the neural network learn how to implement the tasks of programming. This is also a similar work, memory network, learning to manage this long-time memory, in the application of the question and answer system, you can learn some reasoning ability.

The second direction is to focus the attention model (Attention models), dynamically focusing the attention to some details, improve the recognition performance. For example, if you look at the picture, you can generate a sentence based on a picture, which is probably very macroscopic. If we can focus on this mechanism from the introduction to the recognition process, according to the current recognition results, the dynamic step to adjust the focus to the image details, then can generate some more reasonable or more fine expression, such as in the image, focus on a UFO, We can adjust the area of interest in the image to find the UFO, extract its characteristics to identify, to obtain more accurate text description of the image.

The third one is reinforcement learning (reinforcement learning). In the framework of reinforcement learning, there are two parts, some of which are autonomous control units (agents), and some are environment (environment). Autonomous control units are rewarded for maximizing their long-term expected returns by selecting different strategies or behaviors, while the environment receives policy behavior, modifies status, and feeds back rewards. In this enhanced learning framework there are two parts, one part is how to choose these behaviors (policy function), and the other part is how to evaluate those gains (value function) that you can make. This reinforcement learning framework itself has been around for many years, and deep learning is a combination of how to choose the function of the strategy behavior and how to evaluate the expected reward function, which is learned from the deep neural network, such as the moves network in Alphago Weiqi (policy Network) and evaluation networks (value networks).

In a word, from the perspective of research, deep learning is slowly from supervised learning to this interactive learning development; network structure from the first to the network to have a recursive approach, consider the memory, consider the timing of the network, the content from the static input to the dynamic input, In terms of forecasting, it is slowly becoming a step-by-step sequence of predictions from simultaneous predictions. From the development of 2014 and 2015, deep learning is now very simplified thinking is that if there is a relatively new problem, the first thing to do is to describe the problem, to ensure that the input to the final purpose of the process is a differential, and then the most difficult part of the deep neural network to achieve end-to-end learning. Some of the new trends mentioned earlier are largely the same.

Whether it is the public or the media, or the researchers themselves, we may have some different perspectives on deep learning. I personally think that this is a very pure computational problem in the field of computer science, and explores how to abstract the content and structure of these data into a better understanding. I hope that some of the new trends of deep learning mentioned today will help and learn from them. Thank you!

Deep learning moves from being supervised to interacting