cool hit counter Research Opportunities and Challenges in Artificial Intelligence_Intefrankly

Research Opportunities and Challenges in Artificial Intelligence


His students include Yoshua Bengio, a leading authority on deep learning, Zoubin Ghahramani, and Wu Enda, former chief scientist at Baidu, among others.

Not long ago, Professor Jordan was hired to chair the Anthem Science Think Tank. On the day of the hiring ceremony, he shared his views on AI with various technical students within Anthem. The following is a selection of talks given by Professor Jordan, the grandfather of AI.

Future technologies such as artificial intelligence, which is changing our world. My field of study has changed even more dramatically in the last 30 years. However, our understanding of how these technologies are being built today and how they will evolve in the future is far from adequate. As a result, we are facing a number of technical difficulties and challenges.

To start with a brief introduction, I joined the University of California, Berkeley in 1998 as a professor and have 30 years of research experience in the field of machine learning as well as statistics. As a result, I am very interested in the study of large-scale data.

These big data, which have been growing and expanding in size for the past few years, have been very exciting for us scientists and researchers. Using this rich data, we can create more segmented markets and services.

These imaginative business models and markets, such as the financial markets, are very exciting and thrilling to me. I am therefore very excited to be joining the Anthem team and I am very happy to be working with my colleagues at Anthem.

Some Perspectives on Artificial Intelligence

First of all give I want to discuss with you about ArtificialIntelligence (AI).

More and more people are talking about AI. The rapid growth of artificial intelligence over the last 10-20 years has actually been the growth of machine learning and the development of statistics. But at the end of the day, both the development of AI and the development of machine learning rely on the accumulation and development of data at their core.

Today, the word AI is constantly mentioned in high frequency. So what exactly is AI, and what is its goal? How far has artificial intelligence come now, and what's next?

First of all, the first image that comes to mind when artificial intelligence is mentioned is usually a robot. Like the intelligent robot in the movie I, Robot, humans can interact with it in an intelligent way.

As a result, some people feel that AI means advances in this area. It can communicate with you and even take care of your food, clothing and shelter. It's also a common image of artificial intelligence we see in artwork like movies.

The second common idea about AI, we call it Augmented Intelligence (IA), or IA for short.

What do you mean by augmented intelligence? It's as if you use a search engine to search for the word and you find that it returns to you a huge number of results in a very short period of time, results that you would be completely unable to remember with the power of your human brain. Augmented intelligence is like a tool like a search engine, which can help you perform tasks that were previously beyond human reach.

In addition to search engines, augmented intelligence technologies that you can experience on a daily basis include recommendation systems, where websites can provide you with more personalized recommendations based on your preferences, and machine translation systems, where computers can easily help you switch between multiple languages at your own pace.

The third one is at the AI infrastructure level. On a broader level, AI can moreover represent the infrastructure around people's lives, such as transportation networks, smart homes, urban planning, and even financial networks. Combined with AI technology, these infrastructures can be smarter and more predictable.

We can intuitively feel the impact that AI technology has had on our lives, something that I'm sure we all feel in China as well. We can get more information more easily and use that information to do what we want to do.

Finally, there is a "hybrid" view of AI, which is an integration of the above three views. Such as autonomous driving, artificial intelligence doctor's assistants, and educational platforms. It is not only able to interact with people as an entity, but it also empowers people from different angles and creates greater value.

What is achievable with artificial intelligence and what is not

Of course, we can't implement everything. In the following, I will share with you the current research progress on four research directions: computer vision, speech recognition, natural language processing, and robotics.

computer vision

Much has been explored about the future vision of computer vision. Image recognition technology that may not have been possible a decade or so ago has now achieved a qualitative breakthrough. Computers are now able to accurately identify specific objects in complex images. But computers currently lack a common-sense understanding of visual scenes.

For example, if I approach the edge of the stage, you get the feeling that I'm likely to fall off the stage. You can tell from the scene what will happen next and why the scene is happening now. We are currently far from being able to achieve this, but it is possible in the future.

speech recognition

Where have we progressed so far in research on speech recognition? Currently, the interconversion from speech to text has been successfully applied in many languages. However, the auditory capabilities of computers are currently very limited.

For example, if you close your eyes and use only your sense of hearing to sense your surroundings, you can tell whether you are in a quiet park or a busy street, and you can deduce the orientation of people and objects around you based on the sounds.

From an auditory point of view, computers currently lack this type of common-sense cognition, and it is even more difficult when complex verbal information is added to the mix.

natural language processing

In contrast to the computer vision and speech recognition problems mentioned earlier - which are still relatively easy - natural language processing is very difficult. Of course, we can see that machine translation has made a lot of progress so far, but it still misses many details in the language.

There is no doubt that the neural network technology used in today's machine translation is capable of computing and matching huge amounts of data in different languages. But humans learn language in a very different way than computers do.

For example, I also know Italian, but when I translate Italian into English, I am more interested in understanding and digesting the Italian phrase and then expressing that semantics in English.

moreover, question and answer(QA) also natural language processing The classic problem of research。 Current research in question-and-answer systems can only answer some conditions explicitly、 Short questions with simple answers, And the inability to provide complex answers to complex questions in real-world Q&A scenarios。

Finally, the semantics in people's languages are complex and diverse, with issues such as synonyms, near-synonyms and antonyms, and a single phrase may imply multiple meanings in different linguistic scenarios. Expressions and conventions vary even more between languages.

For humans, we learn how to recognize these complex contexts as we grow up, but computers are currently far from being able to do that.

Robotics Science

The robots currently in use in industry are only programmed to perform a few fixed tasks, which is a far cry from the "artificially intelligent robots" we imagine. Robotics science contributes to the ultimate vision of AI research - a future where we want AI robots to be able to operate autonomously and interact with us.

A vision for the next decade of artificial intelligence

Next, I discuss with you the vision for the next decade of artificial intelligence.

While I am not a prophet, I think that those listed above are not achievable today, but have the potential to become reality in the next decade. There are many companies and institutions in the industry that are engaged in research in these areas with a view to eventually coming up with a suitable solution.

For example, in the next decade, self-driving cars and even driverless air taxis are possible, and although the experience of using these technologies is not yet very good, it can be expected that these cutting-edge technologies should be available to people in the next decade.

On top of the availability of technology, it is believed that a more desirable situation will be reached in ten years. Of course, within the next decade, the "intelligence" of AI systems will be so limited that you won't think they're as smart as humans. I don't think these AI systems will be able to be as flexible and creative as humans are in the next decade.

AI systems tend to be limited to a particular domain and they are very limited in the semantics they can understand. As for what kind of understanding AI systems can generate during human-computer interaction, and whether they can achieve advanced intelligence such as prediction and planning - we are actually very far from that step, and it will take at least decades, if not centuries, for robots to understand humans.

In that case, what else about AI research is going to be difficult to achieve in our lifetimes?

Arguably, creativity and intelligence are still difficult to achieve for AI systems, and the implementation of reasoning and abstraction capabilities seems out of reach.

For example, on social media, from time to time, people create a new word and others can easily understand what the word means in that semantic context without having to read through thousands of sentences to understand it like a computer.

In addition, it is very difficult for an AI system to get it to actively make a long-term plan, whereas humans often set ambitious goals for themselves on their own initiative.

In addition, there are many other limitations to the development of AI technology, which is nowhere near as powerful a learner as a child in the formative years. A child can learn about the world through a small number of pictures and information in books, but it is still difficult for an AI to develop its own "understanding" of the world even after seeing countless pictures and information.

I don't really see any super human AI in the foreseeable future. Of course some people who are not in the field of AI research will advocate for super-human robots to appear later. I don't think that will happen, nor is there a reason for it to happen.

Of course you may also disagree with this view, for example you may feel that computers are much more powerful than human processing power. But the current understanding of "intelligence" is so limited that it is impossible to predict how much computing power will be required to achieve true artificial intelligence.

What we can see now is that the computer is capable of processing large amounts of data, but it is still very limited in its ability to make assumptions, reason, etc. While the computer can recognize these scenes, it cannot understand the role and meaning of the scenes.

Humans currently spend a lot of energy on helping machines understand the real world, but computers are not capable of active learning. Computers and humans are vastly different, not to mention higher levels such as self-awareness.

Not long ago AlphaGo swept through the Go world and made people exclaim about the powerful "intelligence" level of artificial intelligence. But I actually don't think Go is a very difficult game, because these games are right in front of you, and the options are limited at each move. But in real life, the judgments we make and the choices we face often have no boundaries - everything is possible in the world outside your door and the world in your head is wide open.

Go, it's true, requires supercomputers to calculate vast amounts of possibilities, but Go players don't think in terms of machines. Therefore, we don't claim that machines are smarter than humans because they beat them at Go.

But the good news is that the powerful computing power and advanced algorithms of AI are dispersing in a variety of different application scenarios. You may think that machines have reached a high level of intelligence, but that assertion is actually an overstatement.

For the intelligence of brilliant humans, the answers to Go problems are limited and therefore relatively simple, while problems like transportation, finance, and healthcare, which usually have diverse solutions, are the really tricky ones.

What we should worry about regarding artificial intelligence

Artificial intelligence systems look smart, but they're not.

First, the AI system doesn't really understand what he's doing. For example, replacing some words and phrases in the system with other words that have similar pronunciation but opposite semantics does not allow the system to detect anomalies at the level of semantic understanding, as long as it is functioning properly.

Second, AI systems do not know what results will be produced after making a search, or providing data. Artificial intelligence can have very serious consequences if it goes wrong, and that is something people need to consider.

Search engines, where the system returns a variety of search results after you enter your keywords. But with a medical diagnosis, you must provide an effective and viable treatment plan, and if a medical diagnosis is wrong, it can be fatal; in the world of finance, bad decisions can trigger huge financial losses; and in transportation issues, bad decisions can invite unnecessary calamities.

Third, AI may make some jobs disappear, but it will also create new ones. We know that the industrial revolution of a few hundred years ago cost some people their jobs, along with the emergence of more new jobs, but it inevitably took time for people to learn and adapt to the transition.

Finally, there is the question of the use of artificial intelligence. I don't think robots will rule humans in the future, although this scenario often appears in movies, novels, and other artistic creations, so the topic is brought up from time to time. I don't think the question is whether AI technologies are dangerous or not per se, but whether they will be used incorrectly by people with bad intentions. We need to use technology in the right scenarios and with the right people to truly empower the world with technology.

What are the important technologies currently available in artificial intelligence

Next, I share with you a few technical directions for AI research.

First, machine learning

Things like clustering, classification, prediction, dimensionality reduction, optimization, etc. are all worthy directions for research. With large data sets, good algorithms and parallel distributed computing, good results can be achieved.

Second, planning

How do you figure out the best solution to a problem? We can find shortcuts to problems based on search technology to power AI strategies and tactics. This is also an aspect of machine learning.

Third, human-computer interaction, which has been an important topic

Human-computer interaction refers not only to getting the machine to work independently, but also to how to facilitate human-computer interaction more effectively. Research interests include how to get machines to actively learn from humans, crowdsourcing to solve complex problems, and economics and game theory models.

Challenges in machine learning

Above, I've shared with you the challenges of machine learning today in the form of a list, and I think there's still a lot of work to come in this area.

Uncertainty issues. Although deep learning is developing rapidly, there are still many problems that need to be solved. In particular, there are still black box issues that have not yet been fully resolved, where people focus only on the inputs and outputs that ultimately lead to the results, with a lot of uncertainty about the process in between. But in addressing issues such as healthcare, this uncertainty cannot be informative if it is high.

Non-explainable issues. We need a system that can explain the reasons behind the decisions and behavior of machines.

Deep understanding of machine learning, understanding every aspect of machine learning. Currently, we are still very dependent on datasets and are unable to use the small amount of data to do other thought processes such as analogy and inference.

Artificial intelligence systems need to be able to set and plan long-term goals and proactively collect relevant data for analysis.

Artificial intelligence systems need to enable real-time and timely performance and feedback. We can only expect the results to be as quick as possible now.

It is also now a challenge to ensure the robustness of the system and solve the problem for some unanticipated scenarios.

The problem of how to ensure robustness when the system faces adversarial attacks.

Data sharing issues. For machine learning, the size of the data and the quality of the data are important. If individuals and institutions can share data, put together and integrate disparate data, this can lead to better results.

Privacy protection issues. I believe this is an important challenge for machine learning, which is a different perspective but relevant to all of your work.

Personalization and Machine Learning

We've seen more and more Silicon Valley companies offering personalized services in recent years, and I believe that's where the trend is going. But to provide these services, we need to get a lot of data from consumers and then let the computers learn and make decisions.

While there are hundreds of companies that are attempting to offer personalized services, none of them actually stand out at the moment. Why does this happen? Let's imagine a scenario like the following.

Mentor (owner): we need a system to provide personalized and intelligent services to replace traditional ones.

STUDENT (STAFF): okay, then I will use the algorithms of these machines while needing so much user data and so on to provide personalized services. To do so, we need to build more architecture, we need more servers, and we need to make those servers serve more models.

Mentor (boss): well, in order to improve our service, we also need to build an artificial intelligence system service, we also need to serve more users ......

Finally, as a company grows in size, the quality of personalized service decreases as a company that originally served 20 users serves 10,000, or even millions, of users.

As a company decision maker you need to consider many factors to provide better service while controlling costs. And these conflicts are difficult to resolve in a short period of time.

In the study of machine learning and statistics, we need to have the concept of time budget. For example, if you search for a keyword, you expect the system to have to return the answer quickly within a few seconds.

And with the current personalized service system, which may have thousands of models running at the same time, this system is very complex. As you get more data, or load more models, in order to retain users it has to get faster, it has to get more and more accurate, but this requirement is the opposite of reality.

In fact, because of the increasing volume of data, the error rate increases and the data is processed more slowly instead. As a result, correctness and time budgets are sometimes difficult to balance. As the number of customers grows, so will the different needs of the users.

Robustness of artificial intelligence systems is important. Thousands of years ago humans began building bridges and houses, and they also contributed to economic development. As time went on and centuries passed, there were many bridges and buildings that collapsed due to various natural disasters and other reasons.

This is the same for data science, where we need to ensure not only the quality of the system in the moment, but also the stability of the system over a long period of time. We need professional engineers to solve these problems, but at the moment we don't have enough capacity.

Speaker Bio: Michael I. Jordan Professor (Distinguished Professor) in the Departments of Electrical Engineering and Computer Science and Statistics at the University of California, Berkeley, a member of the National Academy of Sciences, the American Academy of Engineering, and the American Academy of Arts and Sciences, and named the Most Influential Scholar in CS by Semantic Scholar in 2016.

1、From machine learning tuning to referencing comes to mind
2、DedeCMS arbitrary user password reset vulnerability
3、The future of IPFS must be born into the sun
4、2018 Artificial Intelligence and Underwater Robotics Summit and Underwater Robotics Target Capture Competition to be Held in Dalian
5、237 people in the city have been entered into a database of pedestrian traffic violations and will be exposed

    已推荐到看一看 和朋友分享想法
    最多200字,当前共 发送