The Purpose of Artificial Intelligence English (To What Extent Has Artificial Intelligence Achieved?

 

1. Artificial intelligence software chat GPT

Thank you for inviting. To solve this problem, we invited Dr. Liu Tieyan, the chief researcher of the machine learning group of Microsoft Research Asia, to bring us the current situation of AI in his eyes, including challenges and opportunities.

2. Artificial Intelligence Movies

The machine learning group of Microsoft Research Asia includes all the main directions of machine learning, and promotes the academic frontiers in the field of machine learning at different levels such as theory, algorithm, and application. The current research focus of the group is deep learning, enhanced learning, distributed machine learning, and graph learning. Its research topics also include sorting learning, computing advertising, and cloud pricing.

3. List of artificial intelligence stock leaders

In the past decade or so, the group has published a large number of high-quality papers in top international conferences and journals, helping Microsoft's product department solve many complex problems, and contributing to the open source community the Microsoft Distributed Machine Learning Toolkit (DMTK) and Microsoft Graph Engine (LightLDA)

4. Artificial intelligence AI

LightGBM and others have received widespread attention. The group is recruiting talented individuals and sincerely invites heroes from all walks of life to join us to compete and connect with us in the AI world. Here is the official answer to the dividing line. To say how far artificial intelligence has developed, let's first take a look at the historical process of artificial intelligence from 1956.

5. Artificial intelligence computing power

From the beginning of the Dartmouth Conference to the present 61 year history, the development process has experienced ups and downs, and we can see several ups and downs. At least we have experienced two so-called winters of artificial intelligence. Each rise of artificial intelligence is due to some advanced technological invention, and every time artificial intelligence encounters its bottleneck, it is also because people's expectations for artificial intelligence technology are too high, exceeding the level that its technology can achieve.

6. Artificial Intelligence English

So the withdrawal of funds from governments, foundations, etc. has resulted in researchers not having enough funds to engage in research. So what stage are we at today? Some people say it's the spring of artificial intelligence, some say it's summer, and some are pessimistic, saying it's autumn. Autumn means that winter is coming soon. As scholars of artificial intelligence, how should we view this matter and what can we do? No matter how many people predict, at least today is still a golden age of artificial intelligence.

7. Artificial Intelligence GPT

Why do you say that? Next, let me show you some of the recent achievements in artificial intelligence, which were indeed unexpected in the past decade. Firstly, let's talk about the breakthrough of artificial intelligence in speech recognition and speech synthesis. Recently, artificial intelligence has achieved remarkable results in speech recognition and speech synthesis. In October 2016, a new breakthrough with an error rate of 5.9% was achieved in a speech recognition result released by Microsoft Research Institute in the United States, This is the first time that artificial intelligence technology has achieved an error rate similar to human speech recognition.

8. Artificial Intelligence and Intelligent Manufacturing

Secondly, in terms of images, AI has also made great progress. For example, in the ImageNet contest of image recognition, computers are used to identify 1000 categories of images in the dataset. In 2015, ResNet, a technology from Microsoft Research Asia, won the championship of ImageNet in 2015, with an error rate of 3.5% and a human error rate of about 5.1%.

9. Ranking of the Best Schools in Artificial Intelligence

So it can be seen that in particular fields and categories, the ability of computer in image recognition has exceeded that of human beings. In 2016, Microsoft Research Asia made persistent efforts and won the championship in object segmentation, a task more difficult than image recognition. In addition to voice and image, AI has also made great progress in natural language

10. Artificial Intelligence Big Model

The picture on the left describes that major companies are constantly improving their own voice machine translation standards and technologies. The picture on the right shows that Microsoft released a new function of Microsoft Translator last December. It supports more than 50 languages and can realize real-time translation of multiple people and languages. For example, everyone may come from different countries. As long as we hold the mobile phone and use this APP, we can communicate with each other.

When you say a sentence or input text, what the other person hears/sees is their native language

The speech, images, and language mentioned earlier all sound like perceptual things. As we all know, in recent times, artificial intelligence has also made breakthroughs in areas that we believe may be difficult for machines to achieve success. For example, the picture on the left depicts the use of artificial intelligence technology to play games. You can see that this brick knocking game, after 120 minutes of training, Artificial intelligence has found a very effective way to score.

As we continue to train this artificial intelligence robot, after 240 minutes, it has reached the level of a so-called hardcore player and discovered some tricks that we usually cannot play on our own

The game on the right is a Go game. Everyone knows that AlphaGo is very popular, using deep reinforcement learning technology. After a very long training time and citing a large amount of data for self play, it ultimately achieved an overwhelming advantage by defeating the then world champion Li Shishi 4-1. On last year's IJCAI, the main developers of AlphaGo made a keynote saying that since defeating Li Shishi, AlphaGo has not stopped, Because it is a self play process, it can continue to train as long as it has sufficient computational time and samples.

So we can also understand why at the beginning of this year, Master returned to everyone's view and was able to win 60 consecutive games against Go masters, because the gap was too big. These things were areas that people previously thought artificial intelligence could not reach, but it was precisely because these computer scientists and artificial intelligence scientists constantly imitated human decision-making processes, such as training value networks and policy networks, It is how to evaluate the winning rate based on the current chess game and decide what to take next, rather than simply using these value networks to effectively prune the search tree, thus completing a very meaningful exploration in limited time. All of these are essential advances in artificial intelligence technology, allowing us to see some different results.

After discussing the glory of artificial intelligence, there are actually many questions that require us to calmly think and answer.

Our theme is to open the door to the research of intelligent computing. I would like to discuss from the perspective of a researcher what we can do, and even more meaningful, what we can do. Artificial intelligence may seem very popular on the surface, but if you unfold this magic tablecloth, you will find that it is riddled with holes, and various very basic problems have not been solved, There are even many methodological aspects of philosophy that may not be accurate from our perspective.

Faced with such a situation, a more meaningful thing may be to calm down and engage in fundamental research that can change the current and future status of artificial intelligence, or even reflect on the roadmap of artificial intelligence development, to see if we should restart a path. These things can make us not only drift with the flow, but also monetize and cheat money, But rather, we leave our own footprints on the true path of artificial intelligence development. After several decades, when everyone reminisces and another person stands on stage to tell the story of artificial intelligence for a hundred years, he will mark a star on that map, which tells your story.

What is behind the previous artificial intelligence phenomena? Speaking of technology, what are the two most popular words nowadays, deep learning and reinforcement learning? Simply put, it is an end-to-end learning, and we don't need some.

Feature engineering, instead of using a very complex and large capacity model to directly fit the input and output, allows the model to explore meaningful intermediate expressions. What is reinforcement learning? Simply put, learning machines constantly interact with the environment autonomously, using long-term benefits to guide current decisions and adjusting the optimality of decisions through continuous interaction with the environment.

The reason why deep learning and enhanced learning can achieve great success nowadays is that models with huge capacity are trained based on huge data and computational complexity. Therefore, their success cannot be separated from hardware systems. This is also why GPUs are so popular nowadays, including cloud computing and multi machine collaboration, which have become essential links for us.

This is the current situation of artificial intelligence. In the face of this situation, we should follow the path specified by everyone, train more GPUs to train some models and compete with others, or reflect on whether this path is right or not, and whether there are any problems. The next thing I want to discuss with everyone is the many issues of artificial intelligence. I have only listed some representatives, but in fact, the problems are far more than these.

The first thing is that today's artificial intelligence technology, especially represented by deep learning, requires a large amount of annotated data to enable us to train an effective model that does not rely too much on human prior knowledge. To learn from scratch, a large number of samples need to provide patterns.

For example, image classification now typically uses tens of millions of images for training; Speech recognition, thousands of hours of annotated speech data; Machine translation is usually trained on tens of millions of bilingual pairs. Such data was unimaginable before, but our age is the age of big data, so with these data, it is possible to conduct in-depth learning and training.

But is this a panacea? In fact, it is impossible or difficult to obtain similar data in many fields, such as healthcare. There are many difficult and complex diseases, and there are few cases around the world. How can we collect big data for this category? Therefore, in this sense, if we can find a way to overcome the need for large annotated data, we can make use of current artificial intelligence technology.

Breaking through the boundaries defined by current data is the key to delving deeper into more fields. The second challenge is the size of the model and the difficulty of model training. As mentioned earlier, deep neural networks have many layers and generally have large parameters, with billions of parameters being commonplace. Faced with such networks, there are at least two difficulties, one of which we often mention.

The problem of gradient reduction and gradient explosion is that when the deep network has many layers, the residual or loss function calculated between the output layer and the label is difficult to be effectively transferred to the bottom layer. Therefore, when using this back-propagation training, the network parameters at the bottom layer are usually not easy to be effectively trained, and perform poorly.

People have invented various means to solve it, such as adding some skip level connections. For example, the ResNet technology invented by Microsoft Research Asia does this. There are many kinds of techniques, but these are actually just techniques to solve problems. In retrospect, whether the original problem itself is necessary needs our reflection.

Furthermore, the model explosion mentioned earlier that billions of parameters are commonplace, and what billions or even tens of billions of parameters mean is that the storage capacity of the model itself is very large. For example, if we use a neural network to create a language model, the training dataset given is the web pages on Clueweb's entire network, which is about the size of a billion web pages.

If you want to use a recurrent neural network to train a language model with such data, a simple calculation will show that the size of the model it needs is about 80G to 100G, which may not sound too large. However, the current mainstream GPU board has a storage capacity of 24G, which is already considered high configuration. In other words, the size of 80G to 100G far exceeds the capacity of a GPU card, so it is necessary to do distributed computing, There are many technical difficulties involved in parallel modeling.

Even if there is a GPU card that can accommodate these 80G or 100G models, it may take hundreds of years to process such a large amount of training data, which sounds quite unreliable. Is it necessary to put ourselves in a memory that is not enough, and the calculation time is also very long, which is also unbearable? This is a question worth considering.

When it comes to large models and annotated data, it is necessary to mention distributed computing. Distributed computing sounds like a relatively mature field because the system field has been studying distributed computing for many years, but when it comes to distributed machine learning, it is different. The purpose of doing distributed computing here is to allow us to use more resources to accommodate larger models, Reduce the computation time to an acceptable level, but we do not want to lose the accuracy of the computation.

For example, it used to take hundreds of years to obtain a very accurate language model, but now there are 100 machines. Although the calculation is fast, the resulting language model cannot be used anymore. This is not worth the loss. When it comes to distributed computing, there are two mainstream methods: synchronous parallel and asynchronous parallel.

What is the parallel method of synchronization? Many machines are divided into a subtask, and after each calculation step, everyone needs to wait for each other, exchange the calculation results, and then move forward. This process can ensure that the entire distributed computing process is controllable, know what is happening, model it mathematically, and theoretically guarantee it.

But its problem is the so-called Cannikin law. As long as there is a slow machine in the cluster, the distributed computing will be dragged down by the machine, and will not get a good speedup ratio. So people start to do asynchronous parallel mode. Asynchronous means that each machine does its own thing, does not wait for each other, and pushes the current model updates trained according to their own data to a server, Update the overall model again.

But at this point, a new problem arises, which is the problem of disordered updates. This process cannot be described by our original mathematical model, and it violates some basic assumptions of optimization technology. For example, when we use the random gradient descent method, it can be proven that the optimization process has convergence when using a continuously decreasing learning rate.

This is because the gradient we add each time is calculated based on the previous calculated model. Once added, the gradient may be old and not calculated based on the previous model. It is not clear whether the optimization can still converge, so although the speed is fast, accuracy may not be guaranteed 。

The fourth one, I call it 'tuning black technology', which is a very interesting thing. I attended a forum a while ago and a guest's sentence left a deep impression on me. He said, 'Do you know why many companies now have deep learning laboratories? I haven't heard of a support vector machine laboratory before, why?'? This is because the technical training process like SVM is very simple, and there are few hyperparameter that need to be adjusted. Basically, as long as you do it step by step, the results are almost the same.

But in the matter of deep learning, if we don't use some parameter tuning black technology, we won't get the desired results. The so-called deep learning laboratory is a group of people who can tune parameters. Without them, deep learning won't be as effective. Although it's a joke, there are too many things that need to be adjusted for deep learning ability, such as how to obtain and select training data, how to partition distributed operations, how to design neural network structures, 10 layers Whether it is 100 or 1000 layers, how to connect each layer, what are the rules for model updates, how to set the learning rate, how to aggregate the results of distributed operations from various machines, how to obtain a unified model, and so on. There are too many things that need to be adjusted, and if not adjusted properly in one place, the results may be vastly different.

That's why many results in papers cannot be reproduced. It's not that the paper is definitely wrong, but at least they didn't tell you how to adjust the parameters. They only told you what the model looked like The next challenge, called the black box algorithm, is not only a problem with neural networks, but also a persistent problem in statistical machine learning for many years,就是用一个表达能力很强的黑盒子来拟合想要研究的问题，里面参数很多。

这样一个复杂的黑盒子去做拟合的时候，结果好，皆大欢喜如果结果不好，出现了反例，该怎么解决呢，这里面几亿、几十亿个参数，是谁出了问题呢，其实是非常难排错的事情相反，以前有很多基于逻辑推理的方法，虽然效果没有神经网络好，但是我们知道每一步是为什么做了决策，容易分析、排错。

所以最近几年有一个趋势，就是把基于统计学习的方法和基于符号计算的方法进行结合，造出一个灰盒子，它既具备很强的学习能力，又能在很大程度上是可理解、可支配、可调整的

到现在为止，这几件事都是现在人工智能技术层面的问题接下来，谈的是更像方法论和哲学的问题，仅为个人的观点，跟大家一起分享其中一条，我叫做蛮力解法，舍本逐末这句话什么意思？刚才我提到过深度学习之所以这么成功，是因为它有一个特别强的表达能力，在历史上人们证明过深层神经网络有universal approximation theorem，只要隐结点的数目足够多，任意给一个连续函数，它都可以无限逼近这个函数，换言之，有了很强的表达能力，什么问题都可以学的很好。

听起来好像是挺美好的事，但实际上它背后存在一个问题：它拟合的是数据的表象，数据表象可以非常复杂，但是数据背后驱动的规律是不是真的那么复杂呢，如果我们只看到表象不去研究数据产生的本质，很可能你花了很大的力气去拟合，但是浪费了很多时间，得到的结果也不鲁棒。

举个例子，我们发现大自然也好，人类社会也好，其实没有想象的那么复杂，虽然你看到的很多数据很复杂，它们背后的规律可能非常简单像量子力学有薛定谔方程、量子化学、流体力学、生物遗传学、经济学、社会学也都有类似的简单方程，科学家发现那么纷繁复杂的现象都可以被一个动态系统所刻划，而动态系统背后的规律可能就是一个最高二阶的偏微分方程。

大家可以想象，如果不知道这些机理，不对动态系统做建模，只对动态系统的产出数据做建模，就会觉得这个问题非常复杂，要有一个容量非常大的神经网络去逼近这个数据但反过来，如果目光焦点在这个动态系统身上，可能就两三个参数的一个。

二阶微分方程就搞定了下面也是一个值得思考的问题——动物智能，南辕北辙，虽然前面提到人工智能产生了很多的进步，但其实目前所做的还主要是认知的事情，做一个Pattern Recognition，听听声音，看看图像，这是动物也能做的事。

今天的人工智能没有回答一个关键的问题，就是动物和人的区别可能有人会讲，据说猴子的大脑比人的大脑小很多，有可能是体量的不同但人的祖先跟大猩猩在包容量上应该没有本质的区别，那到底为什么经过漫长的进化，人能成为万物之灵主宰地球了呢？。

我自己的观点是因为人类发明了很多动物界没有的机制和规律比如我们有文字，我们可以把我们对世界的认知，总结出来的规律写下来，把它变成书，变成资料传给我们的下一代当老一辈的人去世之后，孩子们读读书，就掌握了之前几百年几千年人们对世界的认识。

但是老一代大猩猩死掉之后，它的孩子就要从头学起另外，我们人类有强大的教育体系，人从幼儿园开始，小学，中学，一直进入大学，用了十几年的时间，就把几百年、几千年的知识都掌握在身上了，可以站在巨人的肩膀上继续往前走，这非常了不起。

好的老师，会教出好的学生，教学相长，薪火相传这些人类的精髓在今天的人工智能技术里面是并没有充分体现，而没有它们我们怎么能指望深度神经网络达到人的智商呢？前面列了很多方面，是我们对人工智能领域的一些看法，不管是从技术层面，还是方法论层面，都有很多值得进一步挖掘的点，

只有这些问题真正解决了，人工智能才可能稳稳妥妥的往前走，而不只是昙花一现。

基于这些考虑，我所在的微软亚洲研究院机器学习组，对研究方向做了一个相应的布局，比如对偶学习，它解决的就是没有大规模标注数据的时候，该怎么训练一个神经网络、怎么训练一个增强学习模型该论文发表在去年的NIPS大会上，获得了很大的反响。

还有，我们叫精深学习（Light Learning），为什么叫Light？前面提到很多模型太大，放不到GPU里，训练时间很长，我们这个研究就是去回答是否真的需要那么大的模型我们展示了一个很有趣的深度学习算法，叫。

Light RNN，用该技术，只需要用一个非常小的模型在几天之内就可以把整个Clueweb数据学完，而且它得到的结果要比用大模型训练很长时间得到的结果还要好并行学习，之前提到并行学习有很多同步异步之间的权衡，我们发明了一个技术，它有异步并行的效率，但是能达到同步并行的精度，中间的技术解决方案其实很简单，在网上也有论文。

我们用了泰勒展开，一个非常简单的数学工具，把这两者给结合在一起符号学习，就是想去解决黑白之间的那个灰盒子问题自主学习，是想去解决深度学习调参的黑科技，既然调参这么复杂，能不能用另外一个人工智能算法来调呢，能不能用增强学习的方法来调呢，所以我们做了一系列的工作来解决怎么去调各种各样的参数，怎么用另外一个机器学习来做这个机器学习。

最后一个方向，我们叫做超人类学习，我们想受大自然的启发，受人类社会发展的启发，去使得我们的人工智能技术接近人类，甚至超过人类，这背后是整个人工智能方法论的变化如果大家感兴趣，可以关注我们微软亚洲研究院机器学习组，跟我们共同从事机器学习的基础研究。

——这里是回答结束的分割线——感谢大家的阅读本账号为微软亚洲研究院的官方知乎账号本账号立足于计算机领域，特别是人工智能相关的前沿研究，旨在为人工智能的相关研究提供范例，从专业的角度促进公众对人工智能的理解，并为研究人员提供讨论和参与的开放平台，从而共建计算机领域的未来。

微软亚洲研究院的每一位专家都是我们的智囊团，你在这个账号可以阅读到来自计算机科学领域各个不同方向的专家们的见解请大家不要吝惜手里的“邀请”，让我们在分享中共同进步也欢迎大家关注我们的微博和微信账号（搜索：微软亚洲研究院AI头条），了解更多我们研究。